media_intelligence

Flutter bindings for Apple's MediaIntelligence framework (iOS 27 / macOS 27), built on Pigeon platform channels.

It covers two things: grouping the faces in a photo library into people, and finding the highlights and key frame of a video. Both run on the device.

import 'package:media_intelligence/media_intelligence.dart';

final video = await VideoAnalyzer.analyze(
  const MediaAsset(id: 'trip', path: '/path/trip.mov'),
);
video.highlights?.highlights; // the spans worth showing
video.keyFrame;               // a good thumbnail time

final faces = await FaceGroupAnalyzer.open('$appSupport/faces');
await faces.insertOrUpdate(photos).drain<void>();
await faces.update();
final people = await faces.assetIdsByEntity(); // person → photo ids
await faces.dispose();

Requirements

Everything	iOS 27+ / macOS 27+
Builds for	iOS 15+ and macOS 12+
Toolchain	Xcode 27, Flutter 3.38+

On older systems MediaIntelligence.isSupported() returns false, and calls throw MediaIntelligenceErrorCode.unsupported.

The framework reads files. For Photos library items, export them to files first; this package has no Photos integration.

Tested on macOS 27:

• 14 unit tests pass.
• 10 integration tests pass against the real framework. They cover video highlights, levels and key frames on a generated two-scene clip, and the whole face-library lifecycle.
• The example app test passes.
• The iOS example builds.

Not tested: grouping real faces into people. There are no photos of people in this repository. The integration suite has a test for your own photos, which is skipped unless you pass --dart-define=MEDIA_INTELLIGENCE_FACE_DIR=<folder of photos of people>.

iPhone verification (iOS 27): nine integration tests pass, one fails, and the optional real-face test is skipped. The synthetic two-scene clip returns no highlights from Apple's framework on the phone, although the same fixture passes on macOS. Highlight selection still needs validation with representative real video before it can be considered device-verified.

Video

VideoAnalyzer.analyze runs highlight analysis, key frame analysis, or both. Both in one call is cheaper than two calls. Each analysis succeeds or fails on its own, so check highlightsError and keyFrameError.

• highlights.highlights: the spans worth showing.
• highlights.levels: a score for each span, covering the whole video in order.
• keyFrame: the time of the most representative frame.

The first analysis in a process is slow while models load: about 20 s for the 6 s test clip, then about 1–2 s after that.

Face groups

FaceGroupAnalyzer keeps a face library in a working directory you choose. The library persists; open the same directory again on the next launch.

1. insertOrUpdate(assets) adds images, or re-analyzes ids already present. It streams the faces found in each asset. Work starts when you listen.
2. update() regroups faces into people ("entities"). state() reports stale when assets changed since the last update.
3. Query the results:
   • entityIds(), faces(), facesWithIds(...);
   • assetIdsByEntity(), facesByEntity();
   • assetIdsForEntities(...), facesForEntities(...), facesInAssets(...).
4. identifyFaces(assets) matches faces in new images against known people without adding the images.
5. Remove assets with deleteAssets(ids) and deleteAllAssets(). Delete the whole library with FaceGroupAnalyzer.purge(directory).

Asset ids are yours: use a stable key such as a database id. Ids in one call must be unique, and every file must exist. This package checks both before calling the framework, whose own error for a missing file is unspecific.

Face.bounds is the CGRect Apple reports, passed through unchanged. Its coordinate space isn't documented, and this package couldn't observe it without photos of people.

Swift → Dart

MediaIntelligence (Swift)	media_intelligence (Dart)
MediaIntelligenceImageAsset, MediaIntelligenceVideoAsset (.url)	MediaAsset
FaceGroupAnalyzer(workingDirectory:), purge	FaceGroupAnalyzer.open, FaceGroupAnalyzer.purge
insertOrUpdateAssets, identifyFaces	insertOrUpdate, identifyFaces (Streams)
update, state	update(), state()
allEntities, allAssetIDs, allFaces	entityIds(), assetIds(), faces()
allAssetIDsByEntityID, allFacesByEntityID	assetIdsByEntity(), facesByEntity()
fetchFaces(_:), fetchFaces(for:), fetchFaces(in:), fetchAssetIDs(for:)	facesWithIds, facesForEntities, facesInAssets, assetIdsForEntities
deleteAssets, deleteAllAssets	deleteAssets, deleteAllAssets
VideoAnalyzer.shared.analyze(_:for:)	VideoAnalyzer.analyze
HighlightAnalysisRequest.Result, KeyFrameAnalysisRequest.Result	HighlightAnalysis, VideoAnalysis.keyFrame
MediaIntelligenceError	MediaIntelligenceException

Not bridged

• Progress for update(subprogress:). update() only reports completion.

Known platform behavior

On the iOS 27 test iPhone (2026-09-23), the synthetic six-second video returns no highlights and twelve half-second intervals scored zero. Native logging confirmed this is Apple's HighlightAnalysisRequest.Result before Pigeon conversion. The integration test's nonempty-highlight assertion therefore fails on that phone; it remains an open validation issue. The other nine device tests pass, and the optional real-face-photo test is skipped. A representative real video is still needed to validate useful highlight selection on iOS.

Observed on macOS 27:

• Video analysis can't be cancelled. Cancelling the Swift task doesn't stop it.
• A file that isn't a readable video fails each analysis in the result (NSOSStatusErrorDomain -18, "Analysis failed to complete"). It doesn't throw.
• An image without faces is processed, but isn't listed by assetIds().
• A file that isn't an image, such as a video, fails the whole insertOrUpdate call with faceGroupProcessing.
• Cancelling an insertOrUpdate subscription stops the events. The framework may still finish storing the assets.
• purge on a directory that was never opened throws workingDirectory. After a purge the directory itself remains.
• The framework prints progress lines ("Processed asset …") to standard output.

Errors

Every failure is a MediaIntelligenceException with a MediaIntelligenceErrorCode:

• unsupported, invalidArgument, invalidHandle, notFound;
• workingDirectory, mediaProcessing, faceGroupProcessing and resultFetching, which are Apple's MediaIntelligenceError cases;
• mediaIntelligenceError, unknown.

Resources

Call dispose() on a FaceGroupAnalyzer; the library stays on disk. A Finalizer releases forgotten analyzers eventually. After a hot restart, call MediaIntelligence.releaseAll(). MediaIntelligence.liveHandleCount() helps check for leaks.

Development

dart run pigeon --input pigeons/media_intelligence_api.dart
flutter test
cd example
flutter test integration_test/media_intelligence_test.dart -d macos
flutter test integration_test/app_test.dart -d macos

# Regenerate the sample clip and image:
swiftc -O ../tool/make_assets.swift -o /tmp/make_assets && /tmp/make_assets assets

# On a physical device (wireless debugging needs flutter drive):
flutter drive --driver=test_driver/integration_test.dart \
  --target=integration_test/media_intelligence_test.dart -d <device id> \
  --publish-port

Run integration test files one at a time on macOS. package:media_intelligence/testing.dart lets you replace MediaIntelligenceBindings.instance with fakes.