How to Keep AI Characters Consistent Across Every Shot

Story Elements hold three kinds of entity: characters, places and objects. That is your world. It is a canvas, not a form: zoom, pan, move and resize cards, and group them the way you think, main characters near the top, background people further out. Zoom to fit pulls everything back into view. Along the bottom you can add text, images and audio, plus the three that matter most, because the storyboard uses them: character, place, object.

This is not a wiki. Everything you type in a description is read, literally, by the system that draws your images. So test every sentence: would it change the picture? If yes, it belongs here. If not, it belongs in your head, where it is still useful but costs nothing.

Here is what most people type the first time:

Not one of those is an instruction. Pretty is not a look. A coat is not a coat. Worried is an emotion, and emotion changes shot to shot, so it cannot be part of who she permanently is. Generate a sheet from that and you get a different woman every time, because nothing has been decided.

The rewritten description for Mara decides things a camera can see:

• Age and a specific build.
• Hair described by how it behaves, not by an adjective.
• One garment precise enough to redraw: a man's oilskin coat, too big in the shoulders, the cuffs turned back twice.
• A permanent physical detail: a burn scar on the back of her left hand.
• Posture, because posture is character and it survives every camera angle.

There is no backstory in it. Words carry a lot and images carry the rest: drop references into the mood board. Three or four good references beat fifteen. A wall of images that disagree gets averaged, and averaged faces are exactly the bland faces everyone complains about.

In the storyboard stage, every character card has a button that turns the description into a reference sheet: several angles of the same person, from your words and your references. From then on, every shot that includes Mara is generated against that sheet. It is not magic and not memory. It is a picture the system looks at every single time.

Open the advanced settings on any entity to see it plainly: the mood board, your description and your images feed the node that generates the sheet, and the prompt points straight back to your text. Two practical rules:

• If a card shows "primary output missing" in red, the sheet did not generate. Any shot with that entity is now guessing. Regenerate it before you storyboard.
• Changing a description does not update the existing sheet. Use the regenerate button on the card. At 14 credits it is far cheaper than ten shots built on an out-of-date face.

1. 1Shape: circular, concrete, about 8 m across.
2. 2Materials: bare concrete, salt bloom on the walls, steel.
3. 3Light: one source, a tall window facing north over grey water, so everything is cold and comes from one side.
4. 4One distinctive feature: a wall of dead dials with cracked glass.
5. 5State of decay: rust weeping from the door seam.
6. 6What must never change: the sealed steel door with the handwheel is always at the back of the room, opposite the window.

That last line does an enormous amount of work. Without it, the door moves between shots. With it, the geography of your film holds.

The handset: Bakelite, black, heavier than it looks, a coiled cord that has lost its spring, and one detail visible in a close-up, an earpiece worn pale where a thumb has rubbed it for years. That detail is why the object is in the film. It says someone sat here for a long time before Mara did, without a word of dialogue.

In the brief: the coat is a man's oilskin, too big in the shoulders; the window faces north; the door has a handwheel; the earpiece is worn pale. A camera can see every one of those. In your head: whose coat it was, why she took the posting, what happened to the last operator, what the door is for.

The second list is why the first one is good. You only know it is a man's coat because you know whose it was. But typed into the box, backstory dilutes the instruction and the model tries to draw a feeling instead of a coat. Write down what can be photographed.

Type @ in a description to reference another entity's sheet directly, so a shot can say "the empty coat" and the system knows exactly which coat. Entities can point at each other.

The assistant is genuinely good at this if you limit it to what a camera can see. Without that limit you get parents, childhood trauma and a favourite song, none of which will ever appear in a frame. With it, it asks useful questions, like footwear and whether she wears anything on her wrists. It can also cut an over-written description down, and check naming: SIGNAL called the prop a receiver in one action line and a handset in another. Nothing to a human reader, a real problem for a generator.

One character, one place and one object took about 12 minutes. Build the entities that appear in the scene you are about to storyboard, and nothing else. Story Elements is not a task to complete; it is something you feed as you go. Projects die at this stage when someone decides to fill everything in first.

The proof: the same shot and the same prompt, once with "a young woman, pretty, wearing a coat" and once with the real description. That is not a better model. It is a better brief. Next, Lesson 5 turns the scene into a storyboard.