The bottleneck was the PNG encoder — a 16× faster augmentation

Running the Image Processing Tool's default augmentation (factor 5, random ±5°) on 2448×2048 PNGs took about 4.3 seconds per image. Preparing a few thousand training images took hours.

Where the time went

Profiling showed the bottleneck was not rotation or alignment but saving. More than 90% of the run time was spent in PIL's PNG encoder at its default compression level 6.

What changed

1. PNG compression level 1 — still lossless, about 5× faster to encode, roughly 10% larger files.
2. Parallel output saving — the six outputs per input are written concurrently with a ThreadPoolExecutor (up to 8 threads). The PIL encoder releases the GIL, so it scales with cores.
3. Simpler rotation — rotate(expand=True) followed by a center crop became a single rotate(expand=False). Same result, half the cost.
4. Log throttling — per-file logs now print once every 10 files.

Result

Fewer than 0.003% of pixels differ at the rotation boundary, with no effect on augmentation quality.