SAM: Pushing the Limits of Saliency Prediction Models