AI video production: from brief to export

A commercial clip made with AI moves through the same stations as one shot on set. Brief, moodboard, references, keyframes, animation, edit, sound, export. What changes is the cost of trying again. This lesson maps the chain and the models behind each station. It also gives you five questions to judge a shot. Does it read? Does it hold the brand? Does it cut, and does it sound right? Would the client sign it?

Consistency is the whole job. Decide the lens, the depth of field and the light before you write a single prompt. Then keep the same words in every one. Lenses and light are vocabulary the models understand. A moody brief becomes 35mm, shallow focus, single warm key from the left. Write the shot list with a text model, one line per shot, in that vocabulary.

A moodboard is not decoration. It is the set of files you will hand to the model over and over. A character reference, a product reference, a palette, three texture stills. Generate the character once, front-facing, evenly lit, neutral expression. Save it. Every later shot starts from that file, not from a fresh description of the same person.

Animation tools are only as good as the still they start from. For each shot in the list, generate the first frame as an image. Composition, light and product placement locked. Fix small faults in the image editor before animating. An error in the frame becomes an error in every frame. Line the keyframes up in order. If the sequence tells the story as stills, the film will too.

Image-to-video keeps what you approved and adds motion. Text-to-video invents both. Start from the keyframe and describe only the movement. What the camera does, what the subject does, nothing about the look. Simple instructions are more stable than clever ones. Test three motion variants at 4 seconds before you commit to a long take. A change of daylight or a morph inside the shot is a prompt line, not a new shot.

Two techniques carry a character between shots. A start frame and an end frame pin the shot at both ends. The model fills the middle instead of wandering. A reference image keeps the face. For dialogue, generate the line as audio first, then lip-sync the shot to it. An actor reference, a short recording of your own face, steers expression better than any adjective.

Every generative model has the same three weak spots. Hands touching objects, more than two people in frame, precise choreography. Plan around them instead of fighting them. Shoot the product hero without hands, cut to a hand in a separate shot, keep crowds in the background. Then select for story, not for spectacle. A plain shot that cuts well beats a beautiful one that does not.

Video-to-video rewrites what is in the frame without reshooting. Swap the background, remove a passer-by, change the outfit. The trap is scope: the more of the frame you change, the more the rest drifts. Name what must stay, test four seconds, then run the full clip. When more than half the frame changes, a fresh generation is cheaper than a rescue.

Light can move after the fact. A relight pass turns a flat midday shot into golden hour, with shadows that follow. Do it before the upscale, because the upscaler sharpens whatever it is given, faults included. Upscale only the shots that survived selection, at the resolution the platform actually shows. 1080p for social, 4K for a screen or a print frame.

The edit is where AI footage becomes a film. Cut to a tempo, even without music, and hold the hero shot longest. Grade every shot toward one palette so the sequence feels like a single camera. Type and logos go on last, in a still that does not move behind them. Then export the variants the platforms want: 16:9, 9:16, 4:3, 1:1, each reframed rather than cropped blindly.

Sound is half the film and the cheapest half to get right. Generate ambience and effects to the picture, not the other way around. For voice, clone a consenting speaker once and reuse the voice across versions. Translate the script, generate the new track, lip-sync the on-camera lines. One clip becomes five language versions from the same edit.

Footage generated with paid credits is yours to use commercially, within the model's license, but the people in it are not. A face and a voice need a license that names the use and the contexts that are off limits. The EU AI Act asks you to disclose synthetic media where a viewer could be misled. Split responsibility in writing: who generates, who approves, who publishes. Then watch the model landscape, because the tool you learned this month will be replaced next month.