MyLiteRTExperiment

A modern Android app built with Kotlin 2.0.x, Jetpack Compose, and the MVVM architecture, leveraging multiple on-device ML models using the new LiteRT Next (CompiledModel) APIs. The app overlays results from all models on a live CameraX preview, providing a fast, future-proof, and fully on-device AI experience.

Features

• Live CameraX Preview with real-time ML overlays
• Image Classification (MobileNetV3, ImageNet labels)
• Style Transfer
• OCR (Optical Character Recognition)
• Speech-to-Text
• Pure Compose UI with Material 3 theming
• MVVM architecture with repository and model executor layers
• All models use LiteRT Next (CompiledModel) — no classic TFLite code

Requirements

• Android Studio Giraffe or newer
• Kotlin 2.0.x
• Jetpack Compose BOM alpha (2025.06.00 or newer)
• Gradle 8.14.2
• Android device with CameraX and ML support

Setup Instructions

1. Clone the repository:

   git clone <your-repo-url>
   cd MyLiteRTExperiment

2. Open in Android Studio.
3. Sync Gradle and build the project.
4. Run on a real device (emulator support may be limited for CameraX and on-device ML).

Model and Asset Files

• Place all model files (.tflite) in app/src/main/assets/.
• For image classification, use the standard MobileNetV3 float32 model with 1001 output classes (see TensorFlow Lite hosted models).
• Place the ImageNetLabels.txt file (with 1001 labels) in the same assets directory.

Model Compatibility

• Always use models compatible with LiteRT Next.
• Check model input/output shapes and types using Netron:
  • Input: float32[1,224,224,3] (for MobileNetV3)
  • Output: float32[1,1001]
• Do not use custom or non-standard models (e.g., with 1000 outputs) unless you update all code and labels accordingly. Incompatible models may cause native crashes.

Troubleshooting

• Native crash during inference?
  • Check that your model file is not corrupt and matches the expected input/output shapes and types.
  • Use only official, standard models for best compatibility.
  • Ensure only one inference runs at a time per model instance.
• Camera not working?
  • Make sure camera permissions are granted.
  • Test on a real device.
• Gradle or Compose errors?
  • Ensure you are using the correct versions as listed above.

Contributing

Contributions are welcome! Please open issues or pull requests for bug fixes, improvements, or new features. For major changes, please discuss them in an issue first.

License

This project is licensed under the MIT License. See LICENSE for details.