v0.6.6 - Switch export to LiteRT and make Android input layout-adaptive (#548)

🌟 Summary

🚀 v0.6.6 is a major Android model update: the app moves from legacy TFLite export tooling to LiteRT-based .tflite models, adds smarter Android runtime compatibility for both old and new model layouts, and trims the example app by making Snapdragon QNN/NPU support optional.

📊 Key Changes

• Switched Android official model exports to LiteRT 📦
  The release now uses Ultralytics’ newer format=litert export flow instead of the older TFLite export path. Official Android assets now come from v0.6.6 and use the new _w8a32.tflite naming.

• Android runtime now auto-adapts to different input layouts 🔄
  New LiteRT exports use a different tensor layout than older .tflite files. The Android code now detects this automatically, so it can run:

  • new LiteRT NCHW models
  • older legacy NHWC TFLite models
    without users needing to manually adjust anything.

• Fixed support for more model types on Android ✅
  Segment, semantic segmentation, and pose models exported with LiteRT now load and run correctly. The runtime was updated to recognize the newer LiteRT output tensor naming and shape conventions.

• Official Android assets updated to w8a32 LiteRT ⚡
  The project now ships w8a32 LiteRT models for Android, chosen because they stay small, work with the GPU path, and do not require calibration data during export.

• Export scripts were simplified and modernized 🛠️
  The Android export pipeline now targets LiteRT directly, uses the newer quantize setting, removes legacy conversion workarounds, and depends on ultralytics>=8.4.83.

• Snapdragon QNN runtime is now opt-in in the example app 📉
  The example Android app no longer bundles the heavy Qualcomm QNN/NPU runtime by default. This dramatically reduces app size, while still allowing NPU testing when explicitly enabled.

• Small release and workflow polish ✨

  • Example app build number was bumped for store upload acceptance.
  • Slack release notifications were simplified and improved with clickable PR links.

🎯 Purpose & Impact

• Better future compatibility for Android models 🤝
  Moving to LiteRT aligns the Flutter app with the current Ultralytics export path, making it easier to support newer YOLO models and export improvements going forward.

• Smoother upgrades for existing users 🔐
  Because Android now auto-detects input layout and output conventions, older TFLite models should continue working, while new LiteRT models work out of the box too.

• More reliable support across tasks 🎯
  Users working with segmentation, semantic segmentation, and pose should see better stability and fewer model-loading/runtime issues on Android.

• Smaller downloads for Android app testers 📦
  Making QNN optional cuts the default example app size significantly, which helps with faster installs and easier testing. Users who want Snapdragon NPU acceleration can still enable it explicitly.

• Practical performance-focused default for Android ⚡
  The new w8a32 LiteRT assets were selected as a balanced default: small file size, GPU-friendly, and easier export. For most users, that means a more dependable Android deployment path.

• Cleaner maintenance for developers 👨‍💻
  The export and release flow is now simpler, with less legacy logic and clearer model packaging, which should reduce breakage and make future releases easier to manage.

What's Changed

• Make Snapdragon QNN runtime opt-in in the example app by @glenn-jocher in #543
• Bump example app build number to 0.6.5+12 by @glenn-jocher in #544
• Use quantize export arg in TFLite export examples by @glenn-jocher in #545
• Simplify Slack notification messages by @glenn-jocher in #546
• Linkify PR numbers in publish Slack notifications by @glenn-jocher in #547
• Switch export to LiteRT and make Android input layout-adaptive by @glenn-jocher in #548

Full Changelog: v0.6.5...v0.6.6