How to write scene description that directs a performance, not just a setting

Most scene descriptions written for AI video generation read like estate agent copy. They list the contents of a room, the time of day, the colour of the walls. The model obliges and renders exactly that: a well-lit, correctly appointed space with nobody in it doing anything in particular. The result is competent and inert, and you feel it the moment you drop the clip into your timeline.

The problem is not that your descriptions are wrong. It is that they are answering the wrong question. You are telling the model where the scene is set. What you need to be telling it is what is happening inside the person in the frame, and what the camera thinks about that. Performance in AI film lives almost entirely in the gap between what a character's body is doing and what the context makes us feel about it. Close that gap with description and you have directed a performance, even without a single line of dialogue.

Start with the body, not the background

When you describe a scene, your first sentence should not mention the room. It should name a physical state. Not an emotion, a physical state. Emotions are abstract and models render them as generic expressions. Physical states are concrete and models can approximate them.

Compare these two openings:

```
A woman sits at a kitchen table in a quiet apartment, morning light coming through the window.
```

```
A woman sits at a kitchen table, both hands wrapped around a mug she is not drinking from, eyes on the middle distance past the window, shoulders dropped and still.
```

The first gives the model a location. The second gives it a performance. Hands that are holding but not using. Eyes that are looking but not seeing. Shoulders that have given up holding tension. The model does not need to understand grief to render these things. It only needs to execute the physical instruction, and the audience supplies the feeling.

Use the camera's attitude as emotional information

Once you have the body, bring in the camera not as a technical specification but as a point of view. The camera's distance and angle in your description tells the model something about how the scene wants to be regarded.

```
A woman sits at a kitchen table, both hands wrapped around a mug she is not drinking from, eyes on the middle distance past the window, shoulders dropped and still. Shot from across the table, low angle, the camera does not move. The mug and her hands are in the foreground, soft focus.
```

The static low camera across the table puts the viewer at the table with her, at the same level, waiting. The foreground mug makes the inaction the subject. You have now directed the scene's emotional register through spatial information alone.

Layer in what is not happening

One of the most reliable tools in writing performance description for AI is the negative detail: the thing that is present in the frame but being pointedly ignored, or the action that is conspicuously absent.

```
A phone sits face-up on the table to her right, screen lit with a notification she has not looked at.
```

Now the scene has subtext. The audience sees the phone, registers that she is not looking at it, and does the interpretive work themselves. You did not write "she is avoiding something." You wrote the avoidance.

Write the before and after, not just the moment

Scenes feel flat when they exist only in the present tense. Even a static wide shot gains dimensionality if your description implies a before and an after.

```
A woman sits at a kitchen table, both hands wrapped around a mug she is not drinking from, eyes on the middle distance past the window, shoulders dropped and still. A chair on the opposite side of the table has been pushed back at an angle, as if someone left in a hurry. The phone on the table to her right has a lit screen she has not looked at.
```

The angled chair is the before. The unread notification is the after. The woman in the middle of that story is now acting in a scene with a history and a question, and your model only had to render furniture.

The full description, assembled

```
A woman sits at a kitchen table, both hands wrapped around a mug she is not drinking from, eyes on the middle distance past the window, shoulders dropped and still. A chair on the opposite side of the table has been pushed back at an angle, as if someone left in a hurry. A phone sits face-up on the table to her right, screen lit with a notification she has not looked at. Shot from across the table, low angle, static. The mug and her hands are in the foreground, slightly soft. Morning, overcast, no direct sunlight.
```

Notice the description does not once name a feeling. It does not use the word sad, lonely, resigned, or heartbroken. It names a body, objects, their states, a camera position, and a light condition. That is the whole of it, and it is enough.

The difference you will see

With a bare location description, models tend to generate characters in a default presentational state: upright, faced forward, expression resting near neutral or slightly pleasant. The body is unoccupied. Clips generated from performance descriptions, ones built on physical states, negative details, and implied history, tend to hold character positioning more consistently through the clip and give the editor something to cut on. You get a body in a relationship with its environment rather than a figure placed inside a backdrop.

The test is simple. Watch your clip back with the sound off. If nothing is communicated in the first two seconds without context, rewrite the description from the body outward and generate again. Try it on one scene you have already considered finished.