flutter_vision 3.0.2

YOLO object detection and instance segmentation for Flutter on Android and iOS (YOLOv5 to YOLO26), powered by LiteRT with CPU, GPU and NPU acceleration.

3.0.2

• Fix a crash at startup in Android release builds ("Failed to create an instance of class androidx.work.impl.WorkDatabase"). LiteRT pulled in Google Play AI delivery, and with it WorkManager and Room, which R8 breaks; the plugin does not use it and now excludes it, which also makes apps smaller.
• Tested in release mode on a Pixel 6a: 113 ms per photo on CPU, 36 ms on GPU.

3.0.1

• Shorter package description (pub.dev requires 60 to 180 characters).

3.0.0

Complete rewrite. See MIGRATION.md to upgrade from 2.x.

• iOS support (iOS 15+, Swift Package Manager), with the same results as Android.
• New API: YoloModel.load, detectImage, detectFrame, dispose, typed Detection results (Rect box, label, confidence, polygon) and BoxMapper to draw results on screen.
• Several models can be loaded at the same time.
• The model format is detected from the model itself: YOLOv5/v7, v8, v9, v11, v12 and NMS-free YOLOv10 / YOLO26. modelVersion is gone.
• LiteRT 2.3 (CompiledModel API) with CPU, GPU and NPU acceleration and automatic fallback to CPU.
• Correct camera frames on every device: row padding and pixel strides, NV21, YUV420 and BGRA, and a per-frame rotation.
• detectImage applies the EXIF orientation and decodes large photos at the size the model needs (no more out-of-memory errors).
• Letterbox resizing (like Ultralytics) instead of stretching, class-aware NMS and correct boxes for non-square models.
• Segmentation outlines are traced in Dart: OpenCV was removed, the native libraries are 16 KB page aligned (Android 15+) and apps are tens of MB smaller.
• Models with int8 inputs/outputs and NCHW inputs are supported.
• No more crashes when a model is closed or replaced during a detection, and no more native memory leaks.
• The plugin no longer adds camera or storage permissions to your app.
• The pub.dev package is about 120 KB: the example models stay on GitHub.
• Requires Flutter 3.41+, Android minSdk 24, iOS 15. On iOS the simulator needs an Apple Silicon Mac (LiteRT ships arm64 only); see the README for the EXCLUDED_ARCHS setting needed by flutter build ios --simulator.
• Fixes reported issues: 16 KB page size (#58), app size (#27, #40), live detection not working on some phones (#34, #36, #37), segmentation crash with empty masks (#39), several models at once (#43), portrait photos (#41), NCHW inputs (#37) and iOS support (#8). Asset loading errors now say what to check (#44, #51). uint8-input models (#42) are still not supported.
• Tested with Flutter 3.47.8: unit tests, and integration tests on an Android 16 emulator and on an iOS simulator (Xcode 26.6).

2.0.0

• BREAKING CHANGE: Removed all Tesseract OCR functionality.
• Added support for YOLOv11 (YOLO11).
• Improved tensor shape handling for better model compatibility.
• Enhanced memory management and resource cleanup.
• Better error handling and thread safety.
• Updated to support Gradle 8.x and Dart 3.6.0+.
• Performance optimizations across all models.
• Now support orientation and isAsset parameters

1.1.4

• Resolved the YoloV8 output bug and made updates to the latest version. Please find the latest release on the Ultralytics YOLOv8 0.181 GitHub repository.

1.1.3

• Release of segmentation feature via YOLOv8.
• Updated example code.
• Updated README.

1.1.2

• GPU delegation error has been fixed.
• Coordinate representation of the box in documentation was fixed.
• Switching between models now are supported.
• Added quantization option for more efficient models at the cost of some precision.

1.1.1

• Bounding box error has been fixed.
• Confidence scores for Yolov8 has been fixed.

1.1.0

• loadOcrModel, ocrOnFrame, and closeOcrModel have been removed. Instead, Yolo and Tesseract operate independently of each other.
• Models no longer returns responseHandler as output. Instead, it returns a List<Map<String, dynamic>>.
• The Tesseract model has been updated to version 5.0.0, resulting in improved accuracy.
• New methods have been added: loadYoloModel, yoloOnFrame, yoloOnImage, closeYoloModel, loadTesseractModel, tesseractOnImage, and closeTesseractModel.
• Support is now available for both Yolov5 and Yolov8.
• Resource management has been improved, and all models now operate in the background.

1.0.0

• Methods for Yolov5 now is available (loadYoloModel, yoloOnFrame, closeYoloModel).
• Yolov5 and OCR model now is independent one to each other.

0.0.2

• best parameter has been removed.

0.0.1

• Initial release.