What is affected when a custom "input_shape" is used for a pre-trained model?

I’m wondering how the pre-trained models actually handle the variety of inputs? Do the original layer weights stay exactly the same?

Many thanks in advance for your care & time.