We propose a general approach for modifying image-based models to improve robustness and temporal consistency on video.