From a resolution increase of 1.0mp to 2.5mp, the time cost growth typically far exceeds 2.5×, because the computational complexity of self-attention is proportional to the square of the number of tokens.