flutter_litert 3.2.1

LiteRT (formerly TensorFlow Lite) Flutter plugin. Drop-in on-device ML inference with bundled native libraries for supported native platforms and web runtimes.

example/lib/main.dart

// This example requires bundled assets (the .tflite model, label map, and
// sample images) that live in the example/ directory of the repository.
// If you are copying this file from pub.dev, clone the full repo and run
// the app from the example/ subdirectory so that the assets are available:
//
//   git clone https://github.com/hugocornellier/flutter_litert
//   cd flutter_litert/example
//   flutter run

import 'dart:convert' show utf8;
import 'dart:io' show Platform;
import 'dart:isolate';
import 'dart:math' as math;
import 'dart:typed_data';
import 'dart:ui' as ui;

import 'package:flutter/material.dart';
import 'package:flutter/services.dart';
import 'package:flutter_litert/native.dart';
import 'package:opencv_dart/opencv_dart.dart' as cv;

void main() {
  runApp(
    const MaterialApp(
      debugShowCheckedModeBanner: false,
      home: _DetectionDemo(),
    ),
  );
}

// ─────────────────────────────────────────────────────────────────────────────
// Result type
// ─────────────────────────────────────────────────────────────────────────────

class _Detection {
  final String label;
  final double score;
  // Normalized [0,1] in source-image space: xmin, ymin, xmax, ymax.
  final List<double> box;

  const _Detection({
    required this.label,
    required this.score,
    required this.box,
  });

  factory _Detection.fromMap(Map m) => _Detection(
    label: m['label'] as String,
    score: (m['score'] as num).toDouble(),
    box: List<double>.from(
      (m['box'] as List).cast<num>().map((v) => v.toDouble()),
    ),
  );
}

// ─────────────────────────────────────────────────────────────────────────────
// Isolate startup data
// ─────────────────────────────────────────────────────────────────────────────

class _StartupData {
  final SendPort sendPort;
  final TransferableTypedData modelBytes;
  final TransferableTypedData labelsBytes;

  // Engine selection, flattened to primitives so it crosses the isolate
  // boundary cleanly.
  final String engineKind; // EngineKind.name
  final String performanceModeName; // interpreter
  final int? numThreads; // interpreter
  final List<String> acceleratorNames; // compiledModel
  final String precisionName; // compiledModel
  final bool runAsync; // compiledModel

  _StartupData({
    required this.sendPort,
    required this.modelBytes,
    required this.labelsBytes,
    required this.engineKind,
    required this.performanceModeName,
    required this.numThreads,
    required this.acceleratorNames,
    required this.precisionName,
    required this.runAsync,
  });
}

// ─────────────────────────────────────────────────────────────────────────────
// IsolateWorkerBase subclass: wires up the isolate RPC channel
// ─────────────────────────────────────────────────────────────────────────────

class _DetectorWorker extends IsolateWorkerBase {
  @override
  String get workerDisposeOp => 'dispose';

  Future<void> initialize({
    required Uint8List modelBytes,
    required Uint8List labelsBytes,
    required EngineConfig engine,
  }) async {
    await initWorker(
      (sendPort) => Isolate.spawn(
        _Detector._isolateEntry,
        _StartupData(
          sendPort: sendPort,
          modelBytes: TransferableTypedData.fromList([modelBytes]),
          labelsBytes: TransferableTypedData.fromList([labelsBytes]),
          engineKind: engine.kind.name,
          performanceModeName: engine.perfMode.name,
          numThreads: null,
          acceleratorNames: engine.accelerators
              .map((a) => a.name)
              .toList(growable: false),
          precisionName: engine.precision.name,
          runAsync: engine.runAsync,
        ),
        debugName: 'flutter_litert.detector',
      ),
      timeout: const Duration(seconds: 30),
      timeoutMessage: 'Detector isolate initialization timed out',
    );
  }
}

// ─────────────────────────────────────────────────────────────────────────────
// Public detector: thin wrapper that delegates to the background isolate
// ─────────────────────────────────────────────────────────────────────────────

class _Detector {
  _DetectorWorker? _worker;

  bool get isReady => _worker?.isReady ?? false;

  Future<void> initialize({EngineConfig engine = const EngineConfig()}) async {
    final modelData = await rootBundle.load('assets/efficientdet_lite0.tflite');
    final labelsData = await rootBundle.load('assets/labelmap.txt');
    final worker = _DetectorWorker();
    await worker.initialize(
      modelBytes: modelData.buffer.asUint8List(),
      labelsBytes: labelsData.buffer.asUint8List(),
      engine: engine,
    );
    _worker = worker;
  }

  Future<(List<_Detection>, int)> detect(
    Uint8List imageBytes, {
    double threshold = 0.5,
  }) async {
    if (!isReady) throw StateError('Detector not initialized.');
    final Map<String, dynamic> result = await _worker!
        .sendRequest<Map<String, dynamic>>('detect', {
          'bytes': TransferableTypedData.fromList([imageBytes]),
          'threshold': threshold,
        });
    final detections = (result['detections'] as List)
        .map((m) => _Detection.fromMap(m as Map))
        .toList();
    final inferenceUs = result['inferenceUs'] as int;
    return (detections, inferenceUs);
  }

  Future<void> dispose() async {
    await _worker?.dispose();
    _worker = null;
  }

  // ───────────────────────────────────────────────────────────────────────────
  // Isolate entry: all LiteRT work runs here, never on the UI thread
  // ───────────────────────────────────────────────────────────────────────────

  @pragma('vm:entry-point')
  static void _isolateEntry(_StartupData data) async {
    final SendPort mainSendPort = data.sendPort;
    final ReceivePort workerReceivePort = ReceivePort();

    Interpreter? interpreter;
    Delegate? delegate;
    CompiledModel? compiled;
    List<List<double>>? anchors;
    List<String>? labels;
    int inputW = 0, inputH = 0;
    int boxesIdx = 0, classesIdx = 1, numClasses = 0;

    // Engine-specific inference. Returns (boxes, classes, inferenceUs).
    late Future<(Float32List, Float32List, int)> Function(Float32List tensor)
    runInference;

    try {
      final modelBytes = data.modelBytes.materialize().asUint8List();
      final labelsBytes = data.labelsBytes.materialize().asUint8List();
      labels = utf8
          .decode(labelsBytes, allowMalformed: true)
          .split('\n')
          .map((s) => s.trim())
          .where((s) => s.isNotEmpty)
          .toList(growable: false);

      if (data.engineKind == EngineKind.compiledModel.name) {
        // LiteRT Next CompiledModel path.
        final accelerators = data.acceleratorNames
            .map(Accelerator.values.byName)
            .toSet();
        final precision = Precision.values.byName(data.precisionName);
        final bool useAsync = data.runAsync;
        final cm = CompiledModel.fromBuffer(
          modelBytes,
          accelerators: accelerators,
          precision: precision,
        );
        compiled = cm;

        // CompiledModel exposes byte sizes, not shapes. Derive the geometry:
        // a square H x W x 3 input, the boxes output (anchors x 4 floats), and
        // the classes output (anchors x numClasses floats).
        final int inFloats = cm.inputByteSizes[0] ~/ 4;
        inputW = inputH = math.sqrt(inFloats / 3).round();
        anchors = _generateAnchors(inputW);
        final int n = anchors.length;
        for (int i = 0; i < cm.outputByteSizes.length; i++) {
          final int floats = cm.outputByteSizes[i] ~/ 4;
          if (floats == n * 4) {
            boxesIdx = i;
          } else {
            classesIdx = i;
            numClasses = n == 0 ? 0 : floats ~/ n;
          }
        }

        runInference = (tensor) async {
          final sw = Stopwatch()..start();
          final outs = useAsync
              ? await cm.runAsync([tensor])
              : cm.run([tensor]);
          sw.stop();
          return (outs[boxesIdx], outs[classesIdx], sw.elapsedMicroseconds);
        };
      } else {
        // Classic Interpreter path.
        final performanceMode = PerformanceMode.values.byName(
          data.performanceModeName,
        );
        final perf = PerformanceConfig(
          mode: performanceMode,
          numThreads: data.numThreads,
        );
        final (options, del) = InterpreterFactory.create(perf);
        delegate = del;

        final interp = Interpreter.fromBuffer(modelBytes, options: options);
        interp.allocateTensors();
        interpreter = interp;

        final inputShape = interp.getInputTensor(0).shape;
        inputH = inputShape[1];
        inputW = inputShape[2];

        // Discover which output index is boxes (last dim = 4) vs classes (>4).
        final outputs = interp.getOutputTensors();
        for (int i = 0; i < outputs.length; i++) {
          final s = outputs[i].shape;
          if (s.length == 3) {
            if (s[2] == 4) boxesIdx = i;
            if (s[2] > 4) {
              classesIdx = i;
              numClasses = s[2];
            }
          }
        }

        final v = TensorFloat32Views.capture(interp);
        anchors = _generateAnchors(inputW);

        runInference = (tensor) async {
          final sw = Stopwatch()..start();
          v.inputs[0].setAll(0, tensor);
          interp.invoke();
          sw.stop();
          return (
            v.outputs[boxesIdx],
            v.outputs[classesIdx],
            sw.elapsedMicroseconds,
          );
        };
      }

      mainSendPort.send(workerReceivePort.sendPort);
    } catch (e, st) {
      mainSendPort.send({'error': 'Detector isolate init failed: $e\n$st'});
      return;
    }

    workerReceivePort.listen((message) async {
      if (message is! Map) return;
      final int? id = message['id'] as int?;
      final String? op = message['op'] as String?;
      if (id == null || op == null) return;

      try {
        switch (op) {
          case 'detect':
            final imgBytes = (message['bytes'] as TransferableTypedData)
                .materialize()
                .asUint8List();
            final double threshold = (message['threshold'] as num).toDouble();

            final cv.Mat src = cv.imdecode(imgBytes, cv.IMREAD_COLOR);
            final int srcW = src.cols, srcH = src.rows;
            try {
              // Letterbox-resize to model input dimensions.
              final lb = computeLetterboxParams(
                srcWidth: srcW,
                srcHeight: srcH,
                targetWidth: inputW,
                targetHeight: inputH,
              );
              final cv.Mat resized = cv.resize(src, (
                lb.newWidth,
                lb.newHeight,
              ), interpolation: cv.INTER_LINEAR);
              final cv.Mat padded = cv.copyMakeBorder(
                resized,
                lb.padTop,
                lb.padBottom,
                lb.padLeft,
                lb.padRight,
                cv.BORDER_CONSTANT,
                value: cv.Scalar.black,
              );
              resized.dispose();

              // BGR→RGB + normalize to [-1, 1] (EfficientDet MediaPipe format).
              final Float32List tensor = bgrBytesToSignedFloat32(
                bytes: padded.data,
                totalPixels: inputW * inputH,
              );
              padded.dispose();

              // Run inference on the selected engine.
              final (boxBuf, clsBuf, inferenceUs) = await runInference(tensor);

              // Decode anchors → raw detections in letterboxed model space.
              final raw = _decodeAnchorsAndScore(
                boxBuf: boxBuf,
                clsBuf: clsBuf,
                anchors: anchors!,
                numClasses: numClasses,
                threshold: threshold,
              );

              if (raw.isEmpty) {
                mainSendPort.send({
                  'id': id,
                  'result': {'detections': <Map>[], 'inferenceUs': inferenceUs},
                });
                return;
              }

              // NMS in letterboxed model space.
              final boxes = raw
                  .map((d) => [d.xmin, d.ymin, d.xmax, d.ymax])
                  .toList();
              final scores = raw.map((d) => d.score).toList();
              final kept = weightedNms(
                boxes,
                scores,
                iouThres: 0.45,
                maxDet: 100,
              );

              // Remove letterbox padding and map to source-image [0,1] coords.
              final double pt = lb.padTop / inputH;
              final double pb = lb.padBottom / inputH;
              final double pl = lb.padLeft / inputW;
              final double pr = lb.padRight / inputW;
              final double sx = 1.0 - (pl + pr);
              final double sy = 1.0 - (pt + pb);

              double clamp01(double v) => v.clamp(0.0, 1.0);

              final result = <Map<String, dynamic>>[];
              for (final r in kept) {
                final d = raw[r.index];
                final String name =
                    d.classIdx >= 0 && d.classIdx < labels!.length
                    ? labels[d.classIdx]
                    : '???';
                result.add({
                  'label': name,
                  'score': r.score,
                  'box': [
                    clamp01((r.box[0] - pl) / sx),
                    clamp01((r.box[1] - pt) / sy),
                    clamp01((r.box[2] - pl) / sx),
                    clamp01((r.box[3] - pt) / sy),
                  ],
                });
              }

              mainSendPort.send({
                'id': id,
                'result': {'detections': result, 'inferenceUs': inferenceUs},
              });
            } finally {
              src.dispose();
            }

          case 'dispose':
            interpreter?.close();
            delegate?.delete();
            compiled?.close();
            workerReceivePort.close();
        }
      } catch (e, st) {
        mainSendPort.send({'id': id, 'error': '$e\n$st'});
      }
    });
  }

  // ───────────────────────────────────────────────────────────────────────────
  // Helpers (run inside the isolate)
  // ───────────────────────────────────────────────────────────────────────────

  // EfficientDet anchor generator: FPN levels P3-P7, 9 anchors per location.
  static List<List<double>> _generateAnchors(int imageSize) {
    const int minLevel = 3, maxLevel = 7, numScales = 3;
    const List<double> aspectRatios = [1.0, 2.0, 0.5];
    const double anchorScale = 4.0;
    final anchors = <List<double>>[];
    for (int level = minLevel; level <= maxLevel; level++) {
      final int stride = 1 << level;
      final int featureSize = (imageSize / stride).ceil();
      final double baseSize = anchorScale * stride.toDouble();
      for (int y = 0; y < featureSize; y++) {
        for (int x = 0; x < featureSize; x++) {
          final double cy = (y + 0.5) * stride / imageSize;
          final double cx = (x + 0.5) * stride / imageSize;
          for (int s = 0; s < numScales; s++) {
            final double scale = math.pow(2, s / numScales).toDouble();
            for (final aspect in aspectRatios) {
              final double sqA = math.sqrt(aspect);
              final double w = baseSize * scale * sqA / imageSize;
              final double h = baseSize * scale / sqA / imageSize;
              anchors.add([cx, cy, w, h]);
            }
          }
        }
      }
    }
    return anchors;
  }

  static List<_RawDetection> _decodeAnchorsAndScore({
    required Float32List boxBuf,
    required Float32List clsBuf,
    required List<List<double>> anchors,
    required int numClasses,
    required double threshold,
  }) {
    final int n = anchors.length;
    // Pre-compute the minimum logit that can pass sigmoid(threshold).
    final double minLogit = threshold > 0 && threshold < 1
        ? math.log(threshold / (1.0 - threshold))
        : -1e9;

    final out = <_RawDetection>[];
    for (int i = 0; i < n; i++) {
      final int clsBase = i * numClasses;
      // Find the highest-scoring class for this anchor.
      double bestLogit = -double.infinity;
      int bestCls = 0;
      for (int c = 0; c < numClasses; c++) {
        final double v = clsBuf[clsBase + c];
        if (v > bestLogit) {
          bestLogit = v;
          bestCls = c;
        }
      }
      if (bestLogit < minLogit) continue;
      final double score = sigmoid(bestLogit);
      if (score < threshold) continue;

      // Decode box deltas (RetinaNet / EfficientDet [ty, tx, th, tw] format).
      final List<double> a = anchors[i];
      final double cxA = a[0], cyA = a[1], wA = a[2], hA = a[3];
      final int boxBase = i * 4;
      final double cy = boxBuf[boxBase + 0] * hA + cyA;
      final double cx = boxBuf[boxBase + 1] * wA + cxA;
      final double h = math.exp(boxBuf[boxBase + 2]) * hA;
      final double w = math.exp(boxBuf[boxBase + 3]) * wA;

      final double xmin = (cx - w * 0.5).clamp(0.0, 1.0);
      final double ymin = (cy - h * 0.5).clamp(0.0, 1.0);
      final double xmax = (cx + w * 0.5).clamp(0.0, 1.0);
      final double ymax = (cy + h * 0.5).clamp(0.0, 1.0);
      if (xmax - xmin < 1e-3 || ymax - ymin < 1e-3) continue;

      out.add(
        _RawDetection(
          xmin: xmin,
          ymin: ymin,
          xmax: xmax,
          ymax: ymax,
          score: score,
          classIdx: bestCls,
        ),
      );
    }
    return out;
  }
}

class _RawDetection {
  final double xmin, ymin, xmax, ymax, score;
  final int classIdx;
  const _RawDetection({
    required this.xmin,
    required this.ymin,
    required this.xmax,
    required this.ymax,
    required this.score,
    required this.classIdx,
  });
}

// ─────────────────────────────────────────────────────────────────────────────
// UI
// ─────────────────────────────────────────────────────────────────────────────

// Engine configuration: Interpreter vs CompiledModel, platform-aware.
// Top-level so both the UI and the isolate can read the availability helpers.

enum EngineKind { interpreter, compiledModel }

/// Selected inference engine plus its options.
class EngineConfig {
  final EngineKind kind;

  // Interpreter path.
  final PerformanceMode perfMode;

  // CompiledModel path.
  final Set<Accelerator> accelerators;
  final Precision precision;
  final bool runAsync;

  const EngineConfig({
    this.kind = EngineKind.interpreter,
    this.perfMode = PerformanceMode.auto,
    this.accelerators = const {Accelerator.gpu, Accelerator.cpu},
    this.precision = Precision.fp16,
    this.runAsync = false,
  });

  EngineConfig copyWith({
    EngineKind? kind,
    PerformanceMode? perfMode,
    Set<Accelerator>? accelerators,
    Precision? precision,
    bool? runAsync,
  }) => EngineConfig(
    kind: kind ?? this.kind,
    perfMode: perfMode ?? this.perfMode,
    accelerators: accelerators ?? this.accelerators,
    precision: precision ?? this.precision,
    runAsync: runAsync ?? this.runAsync,
  );

  String get label {
    if (kind == EngineKind.interpreter) {
      return 'Interpreter · ${perfModeLabel(perfMode)}';
    }
    final bool gpu = accelerators.contains(Accelerator.gpu);
    final bool cpu = accelerators.contains(Accelerator.cpu);
    final String acc = gpu ? (cpu ? 'GPU+CPU' : 'GPU') : 'CPU';
    final parts = <String>['CompiledModel', acc];
    if (gpu) parts.add(precision == Precision.fp16 ? 'fp16' : 'fp32');
    if (runAsync) parts.add('async');
    return parts.join(' · ');
  }
}

String perfModeLabel(PerformanceMode m) => switch (m) {
  PerformanceMode.disabled => 'CPU (no delegate)',
  PerformanceMode.xnnpack => 'XNNPACK',
  PerformanceMode.gpu => Platform.isAndroid ? 'GPU (GL/CL)' : 'GPU (Metal)',
  PerformanceMode.coreml => 'CoreML',
  PerformanceMode.auto => 'Auto',
};

/// Whether an Interpreter delegate mode is usable on the current platform.
bool perfModeAvailable(PerformanceMode m) => switch (m) {
  PerformanceMode.disabled ||
  PerformanceMode.xnnpack ||
  PerformanceMode.auto => true,
  PerformanceMode.gpu =>
    Platform.isIOS || Platform.isMacOS || Platform.isAndroid,
  PerformanceMode.coreml => Platform.isIOS || Platform.isMacOS,
};

/// Why a mode is unavailable here, or null if it is available.
String? perfModeReason(PerformanceMode m) {
  if (perfModeAvailable(m)) return null;
  return switch (m) {
    PerformanceMode.gpu => 'GPU delegate: Apple and Android only',
    PerformanceMode.coreml => 'CoreML: iOS and macOS only',
    _ => 'Unavailable on this platform',
  };
}

/// CompiledModel GPU acceleration: Metal on Apple, WebGPU on Linux/Windows.
/// The Android GL/CL accelerator is not bundled yet, so Android is CPU-only.
bool cmGpuAvailable() =>
    Platform.isMacOS ||
    Platform.isIOS ||
    Platform.isLinux ||
    Platform.isWindows;

String cmGpuReason() => Platform.isAndroid
    ? 'CompiledModel GPU not bundled on Android yet'
    : 'GPU not available on this platform';

const _kSamples = <(String, String)>[
  ('Street', 'assets/samples/street.jpg'),
  ('Cat', 'assets/samples/cat.jpg'),
  ('Dog', 'assets/samples/dog.jpg'),
  ('People', 'assets/samples/people.jpg'),
];

class _DetectionDemo extends StatefulWidget {
  const _DetectionDemo();

  @override
  State<_DetectionDemo> createState() => _DetectionDemoState();
}

class _DetectionDemoState extends State<_DetectionDemo> {
  final _Detector _detector = _Detector();

  ui.Image? _decodedImage;
  List<_Detection> _detections = const [];
  int _inferenceUs = 0;
  bool _busy = false;
  String? _error;
  int _sampleIdx = 0;
  double _threshold = 0.6;
  EngineConfig _engine = const EngineConfig();

  @override
  void initState() {
    super.initState();
    _initDetector();
  }

  Future<void> _initDetector() async {
    setState(() => _busy = true);
    try {
      await _detector.initialize(engine: _engine);
      await _runOnSample(_sampleIdx);
    } catch (e) {
      if (mounted) setState(() => _error = e.toString());
    } finally {
      if (mounted) setState(() => _busy = false);
    }
  }

  Future<void> _runOnSample(int idx) async {
    if (!_detector.isReady) return;
    setState(() {
      _busy = true;
      _sampleIdx = idx;
      _error = null;
    });
    try {
      final data = await rootBundle.load(_kSamples[idx].$2);
      final bytes = data.buffer.asUint8List();
      final codec = await ui.instantiateImageCodec(bytes);
      final frame = await codec.getNextFrame();
      final (dets, us) = await _detector.detect(bytes, threshold: _threshold);
      if (!mounted) return;
      setState(() {
        _decodedImage = frame.image;
        _detections = dets;
        _inferenceUs = us;
      });
    } catch (e) {
      if (mounted) setState(() => _error = e.toString());
    } finally {
      if (mounted) setState(() => _busy = false);
    }
  }

  Future<void> _applyEngine(EngineConfig engine) async {
    setState(() {
      _engine = engine;
      _busy = true;
      _error = null;
    });
    try {
      await _detector.dispose();
      await _detector.initialize(engine: engine);
      await _runOnSample(_sampleIdx);
    } catch (e) {
      if (mounted) setState(() => _error = e.toString());
    } finally {
      if (mounted) setState(() => _busy = false);
    }
  }

  Future<void> _openSettings() async {
    final result = await showDialog<EngineConfig>(
      context: context,
      builder: (_) => _EngineSettingsDialog(initial: _engine),
    );
    if (result != null) await _applyEngine(result);
  }

  @override
  void dispose() {
    _detector.dispose();
    super.dispose();
  }

  @override
  Widget build(BuildContext context) {
    return Scaffold(
      backgroundColor: Colors.black,
      appBar: AppBar(
        backgroundColor: Colors.black,
        title: Column(
          mainAxisSize: MainAxisSize.min,
          crossAxisAlignment: CrossAxisAlignment.start,
          children: [
            const Text(
              'flutter_litert · Object Detection',
              style: TextStyle(color: Colors.white, fontSize: 18),
            ),
            Text(
              _engine.label,
              style: const TextStyle(color: Colors.white54, fontSize: 11),
            ),
          ],
        ),
        actions: [
          if (_inferenceUs > 0)
            Padding(
              padding: const EdgeInsets.symmetric(horizontal: 8, vertical: 14),
              child: Text(
                '${_inferenceUs}us',
                style: const TextStyle(color: Colors.white70),
              ),
            ),
          IconButton(
            icon: const Icon(Icons.settings, color: Colors.white),
            tooltip: 'Engine settings',
            onPressed: _busy ? null : _openSettings,
          ),
        ],
      ),
      body: Column(
        children: [
          // Sample selector
          SingleChildScrollView(
            scrollDirection: Axis.horizontal,
            padding: const EdgeInsets.symmetric(horizontal: 12, vertical: 8),
            child: Row(
              children: [
                for (int i = 0; i < _kSamples.length; i++)
                  Padding(
                    padding: const EdgeInsets.only(right: 8),
                    child: ChoiceChip(
                      label: Text(_kSamples[i].$1),
                      selected: _sampleIdx == i,
                      onSelected: _busy ? null : (_) => _runOnSample(i),
                    ),
                  ),
              ],
            ),
          ),
          // Confidence threshold slider
          Padding(
            padding: const EdgeInsets.symmetric(horizontal: 16),
            child: Row(
              children: [
                const Text(
                  'Confidence',
                  style: TextStyle(color: Colors.white70, fontSize: 12),
                ),
                Expanded(
                  child: Slider(
                    value: _threshold,
                    min: 0.1,
                    max: 0.95,
                    divisions: 17,
                    label: '${(_threshold * 100).round()}%',
                    onChanged: _busy
                        ? null
                        : (v) => setState(() => _threshold = v),
                    onChangeEnd: _busy ? null : (_) => _runOnSample(_sampleIdx),
                  ),
                ),
                Text(
                  '${(_threshold * 100).round()}%',
                  style: const TextStyle(color: Colors.white70, fontSize: 12),
                ),
              ],
            ),
          ),
          // Image + overlay
          Expanded(
            child: Center(
              child: _busy && _decodedImage == null
                  ? const CircularProgressIndicator(color: Colors.white)
                  : _error != null && _decodedImage == null
                  ? Text(
                      _error!,
                      style: const TextStyle(color: Colors.red),
                      textAlign: TextAlign.center,
                    )
                  : _decodedImage != null
                  ? LayoutBuilder(
                      builder: (context, constraints) {
                        return CustomPaint(
                          size: constraints.biggest,
                          painter: _OverlayPainter(
                            image: _decodedImage!,
                            detections: _detections,
                            busy: _busy,
                          ),
                        );
                      },
                    )
                  : const SizedBox.shrink(),
            ),
          ),
          // Detections list
          if (_detections.isNotEmpty)
            Container(
              color: Colors.black87,
              padding: const EdgeInsets.all(8),
              height: 80,
              child: ListView.builder(
                scrollDirection: Axis.horizontal,
                itemCount: _detections.length,
                itemBuilder: (context, i) {
                  final d = _detections[i];
                  final color = _detectionColor(i);
                  return Padding(
                    padding: const EdgeInsets.only(right: 8),
                    child: Chip(
                      label: Text(
                        '${d.label} ${(d.score * 100).toStringAsFixed(0)}%',
                        style: const TextStyle(fontSize: 12),
                      ),
                      backgroundColor: color,
                      labelStyle: TextStyle(color: _onColor(color)),
                    ),
                  );
                },
              ),
            ),
        ],
      ),
    );
  }
}

// Engine settings dialog: platform-aware picker for the interpreter delegates
// and CompiledModel accelerators that actually work on this target.

class _EngineSettingsDialog extends StatefulWidget {
  final EngineConfig initial;
  const _EngineSettingsDialog({required this.initial});

  @override
  State<_EngineSettingsDialog> createState() => _EngineSettingsDialogState();
}

class _EngineSettingsDialogState extends State<_EngineSettingsDialog> {
  late EngineConfig _cfg = widget.initial;

  static String _accelKey(Set<Accelerator> s) {
    final sorted = s.toList()..sort((a, b) => a.index.compareTo(b.index));
    return sorted.map((e) => e.name).join('+');
  }

  Widget _heading(String text) => Padding(
    padding: const EdgeInsets.only(top: 4, bottom: 4, left: 8),
    child: Text(
      text,
      style: const TextStyle(fontWeight: FontWeight.bold, fontSize: 12),
    ),
  );

  @override
  Widget build(BuildContext context) {
    final bool gpuOn = _cfg.accelerators.contains(Accelerator.gpu);
    return AlertDialog(
      title: const Text('Inference engine'),
      content: SizedBox(
        width: 380,
        child: SingleChildScrollView(
          child: Column(
            mainAxisSize: MainAxisSize.min,
            crossAxisAlignment: CrossAxisAlignment.start,
            children: [
              SegmentedButton<EngineKind>(
                segments: const [
                  ButtonSegment(
                    value: EngineKind.interpreter,
                    label: Text('Interpreter'),
                  ),
                  ButtonSegment(
                    value: EngineKind.compiledModel,
                    label: Text('CompiledModel'),
                  ),
                ],
                selected: {_cfg.kind},
                onSelectionChanged: (sel) => setState(() {
                  final kind = sel.first;
                  var next = _cfg.copyWith(kind: kind);
                  // CompiledModel GPU is unavailable on some platforms; fall
                  // back to CPU so the selection stays valid.
                  if (kind == EngineKind.compiledModel &&
                      !cmGpuAvailable() &&
                      next.accelerators.contains(Accelerator.gpu)) {
                    next = next.copyWith(accelerators: const {Accelerator.cpu});
                  }
                  _cfg = next;
                }),
              ),
              const Divider(height: 24),
              if (_cfg.kind == EngineKind.interpreter)
                ..._interpreterOptions()
              else
                ..._compiledModelOptions(gpuOn),
            ],
          ),
        ),
      ),
      actions: [
        TextButton(
          onPressed: () => Navigator.pop(context),
          child: const Text('Cancel'),
        ),
        FilledButton(
          onPressed: () => Navigator.pop(context, _cfg),
          child: const Text('Apply'),
        ),
      ],
    );
  }

  List<Widget> _interpreterOptions() => [
    _heading('Delegate'),
    RadioGroup<PerformanceMode>(
      groupValue: _cfg.perfMode,
      onChanged: (v) {
        if (v != null) setState(() => _cfg = _cfg.copyWith(perfMode: v));
      },
      child: Column(
        mainAxisSize: MainAxisSize.min,
        children: [
          for (final m in PerformanceMode.values)
            RadioListTile<PerformanceMode>(
              dense: true,
              value: m,
              enabled: perfModeAvailable(m),
              title: Text(perfModeLabel(m)),
              subtitle: perfModeReason(m) == null
                  ? null
                  : Text(
                      perfModeReason(m)!,
                      style: const TextStyle(fontSize: 11),
                    ),
            ),
        ],
      ),
    ),
  ];

  List<Widget> _compiledModelOptions(bool gpuOn) {
    final bool gpuOk = cmGpuAvailable();
    final String? gpuReason = gpuOk ? null : cmGpuReason();
    final String curKey = _accelKey(_cfg.accelerators);

    Widget accel(String label, Set<Accelerator> value, {String? reason}) =>
        RadioListTile<String>(
          dense: true,
          value: _accelKey(value),
          enabled: reason == null,
          title: Text(label),
          subtitle: reason == null
              ? null
              : Text(reason, style: const TextStyle(fontSize: 11)),
        );

    return [
      _heading('Accelerator'),
      RadioGroup<String>(
        groupValue: curKey,
        onChanged: (key) {
          const map = {
            'cpu': {Accelerator.cpu},
            'gpu': {Accelerator.gpu},
            'cpu+gpu': {Accelerator.cpu, Accelerator.gpu},
          };
          final accels = map[key];
          if (accels != null) {
            setState(() => _cfg = _cfg.copyWith(accelerators: accels));
          }
        },
        child: Column(
          mainAxisSize: MainAxisSize.min,
          children: [
            accel('CPU', const {Accelerator.cpu}),
            accel('GPU + CPU fallback', const {
              Accelerator.gpu,
              Accelerator.cpu,
            }, reason: gpuReason),
            accel('GPU only', const {Accelerator.gpu}, reason: gpuReason),
          ],
        ),
      ),
      const Divider(height: 16),
      _heading('Precision (GPU only)'),
      RadioGroup<Precision>(
        groupValue: _cfg.precision,
        onChanged: (v) {
          if (v != null) setState(() => _cfg = _cfg.copyWith(precision: v));
        },
        child: Column(
          mainAxisSize: MainAxisSize.min,
          children: [
            RadioListTile<Precision>(
              dense: true,
              value: Precision.fp16,
              enabled: gpuOn,
              title: const Text('fp16 (faster)'),
            ),
            RadioListTile<Precision>(
              dense: true,
              value: Precision.fp32,
              enabled: gpuOn,
              title: const Text('fp32 (accurate)'),
            ),
          ],
        ),
      ),
      SwitchListTile(
        dense: true,
        value: _cfg.runAsync,
        title: const Text('Async dispatch (runAsync)'),
        onChanged: (v) => setState(() => _cfg = _cfg.copyWith(runAsync: v)),
      ),
    ];
  }
}

// Per-detection overlay colors, cycled by index so adjacent boxes (and their
// chips) stay visually distinct. Wraps around when detections outnumber the
// palette.
const List<Color> _detectionPalette = [
  Color(0xFF00E5FF), // cyan
  Color(0xFFFF5252), // red
  Color(0xFF69F0AE), // green
  Color(0xFFFFD740), // amber
  Color(0xFFFF4081), // pink
  Color(0xFF7C4DFF), // deep purple
  Color(0xFF40C4FF), // light blue
  Color(0xFFB2FF59), // lime
  Color(0xFFFF6E40), // deep orange
  Color(0xFFEEFF41), // yellow
  Color(0xFF1DE9B6), // teal
  Color(0xFFE040FB), // purple
];

Color _detectionColor(int i) => _detectionPalette[i % _detectionPalette.length];

// Black or white text, whichever reads better on [bg].
Color _onColor(Color bg) =>
    bg.computeLuminance() > 0.5 ? Colors.black : Colors.white;

class _OverlayPainter extends CustomPainter {
  final ui.Image image;
  final List<_Detection> detections;
  final bool busy;

  const _OverlayPainter({
    required this.image,
    required this.detections,
    required this.busy,
  });

  @override
  void paint(Canvas canvas, Size size) {
    final double imgW = image.width.toDouble();
    final double imgH = image.height.toDouble();

    // Fit image into canvas while preserving aspect ratio.
    final double scaleX = size.width / imgW;
    final double scaleY = size.height / imgH;
    final double scale = math.min(scaleX, scaleY);
    final double drawW = imgW * scale;
    final double drawH = imgH * scale;
    final double offsetX = (size.width - drawW) / 2;
    final double offsetY = (size.height - drawH) / 2;

    final Rect dst = Rect.fromLTWH(offsetX, offsetY, drawW, drawH);
    final Rect src = Rect.fromLTWH(0, 0, imgW, imgH);
    canvas.drawImageRect(image, src, dst, Paint());

    if (busy) return;

    final boxPaint = Paint()
      ..style = PaintingStyle.stroke
      ..strokeWidth = 2.0;

    final labelBgPaint = Paint()..style = PaintingStyle.fill;

    for (var i = 0; i < detections.length; i++) {
      final d = detections[i];
      final Color color = _detectionColor(i);
      boxPaint.color = color;
      labelBgPaint.color = color.withAlpha(200);

      // Map normalized [0,1] coords to canvas pixels.
      final double x1 = offsetX + d.box[0] * drawW;
      final double y1 = offsetY + d.box[1] * drawH;
      final double x2 = offsetX + d.box[2] * drawW;
      final double y2 = offsetY + d.box[3] * drawH;

      canvas.drawRect(Rect.fromLTRB(x1, y1, x2, y2), boxPaint);

      final String labelText =
          '${d.label} ${(d.score * 100).toStringAsFixed(0)}%';
      final tp = TextPainter(
        text: TextSpan(
          text: labelText,
          style: TextStyle(
            color: _onColor(color),
            fontSize: 12,
            fontWeight: FontWeight.bold,
          ),
        ),
        textDirection: ui.TextDirection.ltr,
      )..layout();

      final double lx = x1;
      final double ly = y1 - tp.height - 2;
      canvas.drawRect(
        Rect.fromLTWH(lx, ly, tp.width + 4, tp.height + 2),
        labelBgPaint,
      );
      tp.paint(canvas, Offset(lx + 2, ly + 1));
    }
  }

  @override
  bool shouldRepaint(covariant _OverlayPainter old) =>
      old.image != image || old.detections != detections || old.busy != busy;
}