training with resize or not

[alicera]

target size is 224x224
There are different size images on training data

method 1 : directly train the network with parameter -s 224 224

method 2: resize the training data to 224x224 and train the network with parameter -s 224 224

the method 2 accuracy is higher than method 1.
anybody know the reason?

training command on this
https://github.com/huggingface/pytorch-image-models