paper said: Although the model was trained on 512x512 inputs, we have extended it so that it can handle arbitrary resolutions1. how to extend it ?