Smart Auto-Resize — Per Image, Per Axis, Never Distorted
Each image is compared against your target per axis independently.
Wider than target → object-aware crop (saliency, not just faces).
Narrower → seamless canvas extension via directional gradient sampling (no banding).
Exact match on an axis → passthrough for that axis.
Switch to Proportional Fit to never crop — only letterbox/pillarbox.
Set Fill → Transparent to keep extended areas as alpha (saves as PNG).
AI Fill uses LaMa ONNX (Carve/LaMa-ONNX, Apache 2.0) — a 51M-parameter
inpainting model running 100% in-browser via onnxruntime-web (WebGPU → WASM fallback).
First run downloads ~208 MB and caches it in IndexedDB forever.