Right now, I'm working on a project using the DINOv2 model, which has been trained using different types of data augmentation, including Gaussian blur. So from my understanding, it should be robust against small changes in the data that involve blurring. However, someone suggested I use Gaussian blur to fine-tune the model using a different cost function, saying that since it's a different cost function, Gaussian blur would make a difference for training.

So my question is more theoretical than practical: is this likely to change anything? Or would that method of data augmentation not change a thing?

Thanks for the help!