flutter_native_ml 1.2.0

Direct access to on-device ML accelerators (Apple Neural Engine / Core ML on iOS, LiteRT GPU delegate, NNAPI and XNNPACK on Android) for fast native inference.

1.2.0

Added

• Zero-copy camera input. model.startCamera() opens the device camera natively (CameraX on Android, AVFoundation on iOS), resizes every frame into the model's input on the model's own thread and streams only the results to Dart. Includes a live preview texture (NativeCameraPreview), lens and resolution selection, cover / fill / contain resizing, mean / std normalisation presets, frame-rate throttling, pause / resume and automatic stale-frame dropping.
• FlutterNativeML.checkCameraPermission() / requestCameraPermission().
• InferenceResult.frame with the source frame size, rotation and timestamp.
• DeviceCapabilities.raw['cameraAvailable'].

Changed

• Android minSdk is now 23 (required by CameraX 1.5); the plugin declares the CAMERA permission, which apps that never use the camera can strip with tools:node="remove".
• iOS apps using the camera must declare NSCameraUsageDescription; the plugin reports MISSING_USAGE_DESCRIPTION instead of crashing when it is absent.

1.1.0

Fixed

• Android did not compile. The Kotlin plugin imported com.google.ai.edge.litert.* classes that no Maven artifact provides. The plugin now depends on LiteRT (com.google.ai.edge.litert:litert / litert-gpu) and uses its org.tensorflow.lite Interpreter API.
• Android returned the tensor type under dtype while Dart expected dataType, which crashed getSignature().
• Android ran model loading and inference on the UI thread.
• Inference with the GPU delegate now always happens on the thread that created the delegate (a LiteRT requirement); every model has its own worker thread.
• iOS mutated shared state from a concurrent queue and returned some results off the main thread.
• iOS ignored the model's declared input shape and always built 1-D arrays.
• iOS/Android accepted only List<double> inputs; integer lists, typed lists and mixed num lists now work.
• startStream / stopStream were not implemented on Android and only emitted timestamps on iOS.
• dart run flutter_native_ml:ml_builder did not work (the CLI lived in a separate package). It now ships as bin/ml_builder.dart.
• Manifest package attribute removed (deprecated with AGP 8).

Added

• ComputeUnit.cpuAndGpu and ComputeUnit.cpuAndNeuralEngine (NNAPI on Android), with automatic CPU fallback. The compute unit actually used is reported in NativeMLModel.acceleratorUsed and InferenceResult.acceleratorUsed.
• loadModel(filePath: ...) for models stored on the device, numThreads and allowFp16 options.
• iOS: .mlmodel / .mlpackage assets are compiled on device and cached, so a single-file asset is enough (compiled .mlmodelc bundles are directories that Flutter does not bundle as one asset).
• iOS: image inputs/outputs (ImageInput, raw or PNG/JPEG), string, int64, double, dictionary and sequence features; float16 arrays.
• Android: all LiteRT tensor types (float32, int8/uint8/int16/int32/int64, bool, string), quantization parameters in the signature, SignatureDef aliases, dynamic shapes via TensorData(data, shape: [...]), native inference timing.
• Streaming API: model.startStream(), model.pushStreamInput(), model.stopStream() with a bounded native queue that drops stale frames.
• FlutterNativeML.getDeviceCapabilities(), getPlatformVersion(), disposeAll().
• NativeMLException with stable error codes instead of raw PlatformExceptions.
• Typed-list outputs (Float32List, Int32List, ...) plus InferenceResult.doubles(), ints(), argmax() and outputShapes.
• Swift Package Manager support alongside CocoaPods; privacy manifest bundled.
• Kotlin DSL build script compatible with Flutter's built-in Kotlin support.

Changed

• Minimum iOS version is 13.0; Android plugin targets Java 17 (as Flutter requires).
• ModelSignature/TensorInfo gained type, elementCount, hasDynamicShape, isOptional, quantization fields and metadata.
• The static FlutterNativeML.startStream/stopStream helpers are deprecated in favour of the NativeMLModel methods.

1.0.1

• Fixed bugs on android.

1.0.0

• Initial release.