There are bit caveats here: Vision transformers outperform CNNs in HUGE data regimes. On small datasets, CNNs still shine. Also, if you take a CNN with modern

There are bit caveats here: Vision transformers outperform CNNs in HUGE data regimes. On small datasets, CNNs still shine. Also, if you take a CNN with modern tricks, they can be on par with vision transformers, e.g convnext Transformers really dominate when you scale the amount of data to infinity