Why your AI scenes don't match shot to shot: the coverage problem

There's a specific frustration in AI filmmaking that gets described as "the scenes don't match up" and diagnosed as a model problem. It isn't. It's the oldest problem in film grammar, and it has a name.

What is coverage?

Coverage is the set of shots you gather so a scene can be cut together — the wide that establishes the space, the singles on each character, the over-shoulders, the inserts. On a set, coverage is what a director shoots so an editor has options.

The word matters because it names what you're actually generating. You're not making four videos. You're making coverage of one scene, and coverage has rules that individual shots don't.

Those rules exist because an audience reads a cut as continuous space and time. Break them and the viewer doesn't think "that shot was wrong" — they lose their bearings and can't say why.

You're not generating shots. You're generating coverage of a space the model can't see.

That's the whole difficulty in one line. A camera on a set is physically inside a real space, so continuity comes free. Each generation is independent — the model has no memory of the previous shot — so every spatial fact has to be restated or it gets reinvented.

Which continuity rules actually break?

Four, in the order they cost you most.

Where this falls short. Naming the rule doesn't fix the shot. Each one has to be written into the prompt for every generation, and the model may still ignore it. What naming it does is tell you what went wrong, which is the part most creators can't do.

How do you specify eyelines?

By stating where the character is looking in world terms, not frame terms.

"Looking left" is a frame instruction and it flips depending which way the camera faces. "Looking toward the doorway" is a world instruction and it survives a reverse angle. The second is what you want, because it means the same thing from any camera position.

For a two-person conversation, decide once which character is on which side of the space, and write that into every shot in the scene. Not which side of the frame — which side of the room.

MEERA is seated left of frame, facing camera-right toward ARJUN who stands by the window. Her eyeline is slightly above the lens, angled camera-right. Window is screen-left behind her, hard afternoon light from that side.

The pattern to notice: the world description stays fixed and only the camera's relationship to it changes.

There's a better way to hold it fixed than retyping it. The reference library has a Locations section, and a saved location gets a handle you call in every prompt — @apartment-kitchen rather than a paragraph you retype and accidentally vary. The space stops being something you maintain and becomes something you point at.

Characters, elements and styles work the same way, and Creative Studio's video engine takes up to 30 image references per generation inside a 50-asset pool. So a location, a character and a style reference can all sit in one shot alongside the prompt.

How do you keep the light matched?

Write the light as a fact about the room rather than a look for the shot.

A scene has one light situation. Naming its source, direction and quality once, then repeating that unchanged in every shot of the scene, is what keeps the grade from jumping. What breaks it is describing the light differently per shot because you want a different mood — the mood belongs in the performance and the framing, not in relighting the room.

The specific failure to watch for: shots generated hours apart, or on different days, drift because you paraphrased the light description. Same problem as paraphrasing a character description, same fix. Write it once, paste it unchanged — or better, save it as part of a location reference so there's nothing to paraphrase.

What order should you generate in?

Wide first, then singles, then inserts. The same order a set shoots for the same reason.

The wide establishes the space and gives you a reference for everything else. Generate it first and you have something to describe the singles against — where the window is, how far apart the characters stand, what's behind them. Generate a single first and you've made a decision about the space without knowing what the space is.

Where this falls short. Generating the wide first costs you a shot you may not use. On a tight budget that feels wasteful, and it usually still pays for itself in re-rolls avoided on the singles.

What about keyframes?

There's a second mechanism worth knowing, and it's a different kind of control.

In Create mode you can anchor a shot with a start frame and, optionally, an end frame — an uploaded image or an actor reference. When both are supplied the model synthesises the motion, lighting shifts and camera path between them.

For coverage that's useful in one specific way: a start frame taken from the tail of your previous shot gives the next shot a spatial anchor that no amount of prose description can match. You're not telling the model where the room is. You're showing it.

Worth knowing that Edit, Extend and Motion modes inherit aspect ratio from the source clip and hide the selector entirely, so framing decisions happen in the first generation rather than later.

When should you break the rules?

When the break is the point, and only then.

Crossing the line deliberately can disorient an audience on purpose — it's a real technique and it works because the convention exists. The difference between a technique and a mistake is whether you knew you were doing it.

The practical test for vertical drama: episodes run thirty to ninety seconds and most viewers watch muted. A disorienting cut costs you more in a format that short than it would in a feature, because there's no time to recover the viewer's bearings.

More on workflow, continuity and shot planning in Production.