flutter_super_resolution 1.1.1

A Flutter package for performing super-resolution on images using ONNX models.

1.1.1

• Documented where to download compatible models: added an AI Zoo link to the README, pre-filtered for type=upscale and Flutter compatibility.

1.1.0

• Much faster output decoding. The model's output tensor is now read as a zero-copy Float32List view over ONNX Runtime's native buffer (via GetTensorMutableData) instead of OrtValue.value, which allocated a nested List of boxed doubles — hundreds of thousands of heap objects per tile. This removes the dominant post-inference cost and its GC pressure. A guarded fallback to the old .value path is kept for forward-compatibility.
• Seamless symmetric tile blending. Tiles are now cross-faded with a raised-cosine (Hann) window feathered on all overlapping edges. Because the rising and falling ramps form a partition of unity, adjacent tiles blend perfectly — replacing the previous one-sided linear ramp that left a slight discontinuity at trailing seams. Implemented as a running weighted-average, so memory stays at 8 × output bytes (no separate float accumulator).
• Faster tensor packing. Input normalisation uses a precomputed 256-entry uint8→float lookup table instead of a per-pixel division; per-tile edge blend weights are precomputed once per tile rather than per pixel.
• Lower per-tile latency. Dropped the per-tile Future.delayed(Duration.zero) GC yield — runAsync already runs inference on a separate isolate (freeing the UI isolate), and the FFI fast-path no longer produces per-tile garbage to collect. The cached GetTensorMutableData binding is resolved once at init.
• Default maxOutputMemoryMB lowered to 128 (mobile-appropriate; matches the example app and documented ~2048×2048 output budget).

1.0.5

• Fixed tiling seams: tile boundaries now fall in the middle of the overlap region (half-overlap stitching) so each tile contributes only from its well-contextualised interior.
• Fixed edge tiles: they are now zero-padded to tileSize×tileSize before inference, preventing shape-mismatch errors on fixed-input ONNX models.
• Fixed ui.Image memory leak: per-tile images are now dispose()d after being drawn onto the canvas.
• Fixed OrtSessionOptions resource leak: release() is now called in a finally block so native options objects are freed on error too.
• Optimised tensor decoding: channel lists are pre-cast once per row instead of once per pixel, significantly reducing dynamic dispatch in the hot loop.
• Added inputName constructor parameter to support models whose input node is not named 'input'.
• Simplified _prepareInputTensor — width/height are now read directly from the image, removing redundant parameters.
• Fixed example app: replaced non-existent ui.ImageForRenderObject with RawImage; moved upscaler initialisation out of the button handler so the model is loaded once rather than on every tap; improved error reporting via SnackBar.
• Fixed test file: removed scaffold Calculator placeholder; added tests for constructor defaults, parameter validation, and lifecycle.

0.0.1

• Initial release.