This page shows you how to set up your development environment to use MediaPipe Tasks in your Flutter and Dart applications. Each task runs Google's MediaPipe runtime for the platform it runs on, and Google's pinned models are bundled with your app at build time. ## Supported platforms and versions To create Flutter applications with MediaPipe Tasks, your development environment requires the following: - Flutter 3.47 or newer. The latest stable release is recommended. - One or more of the following target platforms: - Android 9 (API level 28) or higher - iOS 15.0 or higher - macOS 14.0 or higher on Apple Silicon - Linux x64 - Windows x64 - Web All tasks run on every platform listed above, except Text Proofreader and Text Summarizer, which are not yet available on the web. This matches the MediaPipe Tasks Web SDK, which doesn't support these two tasks. For per-platform requirements, see [Platform notes](https://developers.google.com/edge/mediapipe/solutions/setup_flutter.md.txt#platform_notes). ## MediaPipe Tasks dependencies MediaPipe Tasks provides five Flutter packages for vision, text, audio, retrieval, and decision tasks. Add only the packages your app uses: each one bundles only the MediaPipe library for its domain. Every package depends on `mediapipe_core`, which holds the shared model store and web runtime settings, and re-exports its types, so you don't need to add `mediapipe_core` yourself. ### Vision tasks The MediaPipe Tasks vision package ([`mediapipe_vision`](https://pub.dev/packages/mediapipe_vision)) contains tasks that process images, video frames, and live camera streams (such as Face Landmarker, Hand Landmarker, Pose Landmarker, Object Detector, and Image Segmenter). See [Prepare vision input data](https://developers.google.com/edge/mediapipe/solutions/prepare_data.md.txt#vision) for supported inputs. ```bash flutter pub add mediapipe_vision ``` ### Text tasks The MediaPipe Tasks text package ([`mediapipe_text`](https://pub.dev/packages/mediapipe_text)) contains tasks that analyze and embed text strings (such as Text Classifier, Language Detector, and Text Embedder), plus Text Proofreader and Text Summarizer on platforms other than the web. ```bash flutter pub add mediapipe_text ``` ### Audio tasks The MediaPipe Tasks audio package ([`mediapipe_audio`](https://pub.dev/packages/mediapipe_audio)) contains tasks that classify audio clips and streams, transcribe speech, and synthesize speech (Audio Classifier, Audio Transcriber, and Text-to-Speech). See [Prepare audio input data](https://developers.google.com/edge/mediapipe/solutions/prepare_data.md.txt#audio) for supported inputs. ```bash flutter pub add mediapipe_audio ``` ### Retrieval tasks The MediaPipe Tasks retrieval package ([`mediapipe_retrieval`](https://pub.dev/packages/mediapipe_retrieval)) provides text, image, and audio embeddings in a shared embedding space and an on-device vector index over them (Universal Embedder and Semantic Retriever). ```bash flutter pub add mediapipe_retrieval ``` ### Decision tasks The MediaPipe Tasks decision package ([`mediapipe_decision`](https://pub.dev/packages/mediapipe_decision)) answers yes-or-no, choice, and score questions about a text (Decision Maker). ```bash flutter pub add mediapipe_decision ``` ## Bundle models Apps bundle the models they use at build time, so nothing is downloaded at run time and the app works offline from its first launch. 1. In your app's `pubspec.yaml`, declare the `assets/mediapipe/` folder and list the models each package uses under `hooks.user_defines`. Each package's `XxxModels.byName` (for example, `VisionModels.byName`) lists the names it accepts. ```yaml flutter: assets: - assets/mediapipe/ hooks: user_defines: mediapipe_vision: models: [face_landmarker, hand_landmarker] mediapipe_audio: models: [yamnet] ``` 2. From your app's root, download and verify the listed models. Run this command again whenever the list changes: ```bash dart run mediapipe_core:bundle_models ``` The command downloads each listed model, checks it against its pinned SHA-256 hash, writes it into `assets/mediapipe/`, and removes models that are no longer listed. 3. In continuous integration, verify that the folder matches the list without changing it: ```bash dart run mediapipe_core:bundle_models --check ``` Then create a task with the bundled model: ```dart import 'dart:typed_data'; import 'package:mediapipe_vision/mediapipe_vision.dart'; Future countFaces(Uint8List rgba, int width, int height) async { final landmarker = await FaceLandmarker.create( FaceLandmarkerOptions(model: VisionModels.faceLandmarker), ); try { final result = await landmarker.detect( VisionImage.fromPixels( pixels: rgba, width: width, height: height, format: VisionPixelFormat.rgba, ), ); return result.faceLandmarks.length; } finally { await landmarker.dispose(); } } ``` If a model isn't bundled, `create` throws a `RuntimeUnavailableException` whose `fix` names the `pubspec.yaml` entry to add. To use your own model, pass `modelPath` or `modelBytes` in the task options instead of `model`. ### Download models at run time Downloading models at run time is opt-in. To let `model:` download models that your app doesn't bundle, set `ModelStore.allowDownloads` before you create any tasks: ```dart ModelStore.allowDownloads = true; ``` Downloaded models are verified against the same pinned hashes and cached on the device. Run-time downloads need network access: see the Android and macOS notes in [Platform notes](https://developers.google.com/edge/mediapipe/solutions/setup_flutter.md.txt#platform_notes). ## Platform notes ### Android - Set `minSdk` to 28 (Android 9) or higher in `android/app/build.gradle(.kts)`. - Tasks run on arm64-v8a and x86_64. On 32-bit devices (armeabi-v7a), the vision and audio tasks and the Text Classifier, Text Embedder, and Language Detector tasks run on the CPU. Decision Maker, the retrieval tasks, EmbeddingGemma, Text Proofreader, Text Summarizer, Audio Transcriber, and Text-to-Speech require a 64-bit device. - On the Android emulator, only the CPU delegate is available. - To download models at run time, add the `INTERNET` permission to `android/app/src/main/AndroidManifest.xml`. Flutter adds it only to debug and profile builds, so release builds need it explicitly: ```xml ``` ### iOS - Set the deployment target to iOS 15.0 or higher. - On the iOS Simulator, only the CPU delegate is available. - For App Store submission, declare the required-reason APIs that the MediaPipe libraries call in your app's `ios/Runner/PrivacyInfo.xcprivacy`. See [App Store privacy manifest](https://github.com/google/flutter-mediapipe/blob/main/doc/privacy_and_licenses.md#app-store-privacy-manifest) for the categories to declare. ### macOS - macOS 14.0 or higher on Apple Silicon is required. - Build the app for Apple Silicon only. Release and profile builds also build an Intel (x86_64) slice by default, which the MediaPipe libraries don't provide, so those builds fail. Add these lines to `macos/Runner/Configs/AppInfo.xcconfig`: ARCHS = arm64 EXCLUDED_ARCHS = x86_64 Alternatively, run `flutter config --enable-macos-arm64-only` to apply the same setting to every app built on that machine. - To download models at run time, a sandboxed app needs the network client entitlement in `macos/Runner/DebugProfile.entitlements` and `macos/Runner/Release.entitlements`: ```xml com.apple.security.network.client ``` ### Linux - The MediaPipe Linux library loads EGL and OpenGL ES, even for CPU inference. On Debian or Ubuntu, install them with: ```bash sudo apt-get install libegl1 libgles2 ``` ### Windows - Tasks run on the CPU. ### Web - By default, the MediaPipe JavaScript and WebAssembly runtimes load from jsDelivr, and each file is checked against a pinned SHA-384 hash before it runs. Bundled models load from your app's own assets. - To host the runtimes yourself (for offline use, an intranet, or a strict Content Security Policy), copy them into your app from its root: ```bash dart run mediapipe_core:web_runtime web/mediapipe ``` Then set the base URL before you create the first task: ```dart MediaPipeWebRuntime.baseUrl = 'mediapipe/'; ``` - If your site sends a Content Security Policy, allow `'self'` and `blob:` in `script-src` and `worker-src`, `'wasm-unsafe-eval'` for WebAssembly, and `'self'` and `https://cdn.jsdelivr.net` (unless self-hosted) in `connect-src`. Add `https://storage.googleapis.com` to `connect-src` only if you download models at run time. - Decision Maker and the retrieval tasks require WebGPU in the browser. ## Samples and resources - [Live gallery](https://google.github.io/flutter-mediapipe/): try every supported task in your browser. - [flutter-mediapipe on GitHub](https://github.com/google/flutter-mediapipe): source, the gallery app, a [runnable sample](https://github.com/google/flutter-mediapipe/tree/main/samples) for every task and running mode, and an [object detection app](https://github.com/google/flutter-mediapipe/tree/main/samples/apps/object_detection) for every platform. ## Troubleshooting For help with technical questions related to MediaPipe, visit the [discussion group](https://groups.google.com/forum/#!forum/mediapipe) or [Stack Overflow](https://stackoverflow.com/questions/tagged/mediapipe) for support from the community. To report bugs or make feature requests for the Flutter packages, [file an issue on GitHub](https://github.com/google/flutter-mediapipe/issues).