seika 0.1.0

Smart inpainting for Flutter. Routes automatically between PatchMatch and LaMa neural based on mask complexity analysis.

0.1.0

• iOS support — seika now runs on iOS 16.0+. The C/FFI PatchMatch engine is bundled as seika.framework via Flutter native assets and loaded correctly on both simulator and physical devices.
• Fixed DynamicLibrary loading on iOS: native assets creates seika.framework/seika, not a flat libseika.dylib. Updated _openLib() accordingly.
• Fixed _openLib() and _lib in PatchMatchEngine — both are now static, making them accessible from static isolate methods. Previously caused a compile-time error.

0.0.1

• PatchMatch engine — C/FFI implementation with zero additional download. Multi-scale pyramid, bidirectional NNF, EM voting, and seam blending. Works offline with no model files required.
• LaMa neural engine — Large Mask inpainting via ONNX Runtime. High-quality results for object removal, watermarks, and text. Model downloaded on demand (~209 MB).
• MI-GAN neural engine — Mobile-native inpainting (ICCV 2023). 7× smaller than LaMa (~80 MB), faster CPU inference, optimized for on-device use.
• Model catalog — SeikaModelInfo.all static registry. Each model is independently downloadable, verifiable via SHA-256, and deletable to reclaim storage.
• SeikaModelManager — Stream-based download with progress events, SHA-256 integrity verification, and atomic file rename on completion.
• Auto routing — SeikaEngine.auto analyzes mask complexity (texture variance + edge coherence) and picks the best engine automatically. Falls back to PatchMatch when no model is downloaded.
• Mask dilation — Neural path automatically expands the mask by 8 px before inference, giving the model more context at boundaries and reducing edge artifacts.
• Soft boundary blending — BFS distance field applied after inference for a smooth, seam-free transition between the filled region and the original image.
• Isolate-safe — PatchMatch runs in a background isolate via compute. Neural inference runs via OrtSession.runAsync. Both are safe to call from the UI thread.