In neural style transfer, the parameters of the pretrained feature-extracting CNN remain frozen and are not updated during optimization. Consequently, the extracted content features of the original content image and the style features of the original style image remain constant. These reference features can be precomputed exactly once before the training loop begins, significantly reducing computational overhead. In contrast, the features of the synthesized image must be dynamically extracted during every training iteration because the synthesized image acts as the optimizable model parameter that is continuously updated via backpropagation.