Transformer attention becomes especially expensive at high image resolutions. We can see how the quadratic and linear costs of attention and non-attention operations play out at higher resolutions…