Comment by nutrientharvest

It's cropping the original image then tokenizing it again with less downsampling, not cropping its internal representation.