Video adds time to multimodal AI. A system has to understand what appears in each frame and what changes across frames: actions, scene cuts, camera motion, speech, captions, and the order of events.

On the generation side, current models can create short clips from text, images, or reference video. They are useful for creative exploration, ads, social clips, and motion concepts. They are a poor fit when the video must be precise, long, factual, or faithful to product details.

This chapter explains how to build practical video-understanding pipelines, control cost with sampling, and make good production decisions about video generation.