Improving Motion in Image-to-Video Models via Adaptive Low-Pass Guidance