Weighing resolution, model scale, and augmentation changes against inference cost.
Increasing input resolution gives the largest accuracy improvement but requires longer training and more VRAM — the highest cost, and the first thing to try.
Scaling up the model (nano → small) roughly doubles parameter count and inference time for a meaningful but smaller gain than resolution — apply after resolution is confirmed.
Augmentation (scale, copy-paste, mosaic, mixup) adds training time but no inference cost — the cheapest lever to pull once resolution and scale are set.
Reducing the NMS IoU threshold is free — no retraining required — and should be tried immediately alongside any other change.