CNNs rely on local convolution kernels stacked layer by layer, have limited receptive fields, and excel at capturing local features such as edges and textures; ViTs are better at modeling global context and long-range dependencies.