What AI video can and cannot do today

AI video turns a sentence, a still image or both into a short clip. Google's Veo 3.1, the model this course tests most, makes clips of 4, 6 or 8 seconds with the sound already in them, and Runway, Kling and Luma also generate in single runs measured in seconds rather than minutes. Plan around that length. The unit a model can handle is the shot, one continuous run of camera, so a film made this way is a set of separate shots cut together afterwards, the way a live-action film is. There are three ways in. Text to video starts from words alone, image to video animates a still you supply, and extension continues a clip you already have.

Inside one shot, results can look expensive. Models are good at light moving across a surface, at water, smoke and fabric in motion, and at slow camera moves on a subject that stays still. They handle one clear action well, such as a cup being set down or a door swinging open. Veo also generates its own sound, adding footsteps or a short spoken line when the prompt asks for one. For a mood film or the cutaways between product shots, today's tools already produce footage that runs in paid campaigns.

Consistency is the weak point. A face drifts over a few seconds, so the person in the last frame can look like a cousin of the person in the first. A logo on a moving cup loses its letterforms, lettering on a sign can wobble between frames, crowds melt into each other, and hands doing anything precise gain or lose fingers. The model also adds things nobody asked for, such as the hand in the spoon test that Reference images, and Veo's own sound tells in full. Across ninety seconds and fifteen shots, keeping one particular person or product the same is still the hardest problem in the field.

The reason is temporal coherence, the need for every frame to agree with the frames around it. An image model makes one picture. A video model has to get every frame of a clip right and keep them consistent with each other, and the longer the clip, the more room there is for a face or a label to slide. That is why drift gets worse with length, and why a three-second shot is usually safer than an eight-second one. Each new model version pushes the limit back a little, but the way things fail has stayed the same.

Working professionals design around those limits instead of fighting them. They settle each frame as a still first, where trying again is cheap and control is good, and then ask the video model only for movement. They name the camera move in the words a cinematographer would use, such as a slow push in or a pan left, because models learned from footage described that way. They board the film in short shots with one clear action each, and they put the cuts where a model would struggle, so a face turns away before it can drift. That also describes a well-cut advertisement. A thirty-second commercial is often fifteen cuts.

Cost shapes the process. The Gemini API bills Veo by the second, at $0.05 a second on the Lite tier at 720p and $0.40 on the Standard tier, and every failed take is paid for at the same rate as a good one. A shot that needs twelve attempts costs twelve times the sticker price. Twenty 4-second drafts on Lite come to $4, while a single 8-second take at 1080p on Standard costs $3.20. So guessing is expensive, and teams have quietly brought back storyboards, shot lists and animatics, where a still for every shot is cut to temporary sound before any video is made.

Generated footage still needs finishing. Clips arrive at awkward sizes and frame rates, with a faint shimmer, and clips from the same prompt differ in exposure and colour, so they have to be matched, graded and given grain before they read as one film. Sound carries more of the result than beginners expect, as the sound lesson later in this module explains, which is why the sound Veo makes is best treated as a guide track for anything a client will see.