A video is not a single picture that moves. It is a sequence of still images — frames — shown in order fast enough that the eye reads them as continuous motion. Typical video runs at roughly 24 to 30 frames per second, so even a short five-second clip contains well over a hundred separate images.
Each frame is the same kind of object the previous chapter dealt with: a rectangular grid of pixels holding numeric color values. So generating a video means generating many pixel grids in sequence rather than one. The refinement idea carries over directly — start from noise, adjust step by step under prompt guidance — but now the adjustment has to happen across the whole sequence at once.
This is where the new problem appears. If each frame were refined in isolation, the model would produce a hundred plausible pictures that happen to share a prompt but not a subject. A person's face would subtly change shape between frames; a shirt's color would shift; a background object would slide. The eye is extremely sensitive to this, so the requirement is not just that each frame looks good, but that consecutive frames describe the same world. That requirement is called temporal consistency, and it is the central challenge of video generation.
The practical consequence is that video models must carry information forward in time. Whatever the model decides about the subject in frame one constrains what it can decide in frame two, and so on down the sequence. The conditioning principle from earlier chapters still holds — the prompt steers the output — but it now steers a trajectory rather than a single point.