Image Dimensions for Fine Tuning

[alexisdrakopoulos]

Hi,

I'm looking to fine-tune SIGLIP2, probably using: ViT-B-16-SigLIP2 ?

I have millions of images so I'd like to reduce the transfer cost / size of the dataset. I currently use webdataset tar format.

If I am training a 256 model, is it fine for me to pre-resize my images to 256x256 with a bit of compression, or should I use a higher dimension? I don't see much in terms of image-preprocessing steps during fine tuning. I feel a little lost!