# Why AI actors overact, and how to direct a natural performance

> AI video models play every feeling at full size. Why AI actors overact, and how to write shot directions that get natural, believable acting.

Published 2026-11-15. Source: https://studio.oneinfer.ai/blog/why-ai-actors-overact

## What overacting looks like in AI video

You have seen it even if you never named it:

- A small smile arrives as a full grin, and stays there.
- "Worried" becomes horror: wide eyes, open mouth, hands to the face.
- Every line is delivered loud, as if to the back row of a theatre.
- The face changes every second or two, from one expression to the next.
- Everyone in the frame reacts at once, so you can't tell who the scene is about.

We saw all of this in our own test films while building Movie, Studio's story-to-film tool. When we read the shot directions behind those films, the cause was plain. One eight-second shot asked a single face for "defensive anger that cracks, then resolve, then tired relief and quiet pride". A character who was only listening was given "tearful pride". A hero's every line was "warm and booming". The video model did exactly what it was told, and that was the problem.

## Why AI actors overact

### The model plays every word at full strength

A human actor reads "furious" and decides how much of it to show. A video model has no such judgement: every adjective in the prompt is something it must make visible, and the surest way to make a feeling visible is to make it big. "Slightly annoyed" helps a little. "Annoyed, frustrated, tense" asks for three things, and you get all three.

### Stacked feelings become a performance of changes

"Anger, then resolve, then relief, then pride" in eight seconds leaves two seconds for each. The model obliges with four distinct expressions, one after another. On screen that reads as mugging, not as an inner journey. Real people's faces change far less often than our descriptions of their feelings do.

### Everyone gets a feeling

When a direction gives every person in the frame an emotion, the people who are listening perform as much as the person speaking. Film acting works the other way: the listener is usually still, and that stillness is what makes the speaker's line land.

### "Never leave a face blank" rules

Prompt guides, and the AI writers that follow them, often say never to reset a character to a neutral face. The aim is good, since dead, expressionless faces look wrong. But applied to every shot it pushes every face to show something all the time, and the sum is a film with no quiet moments.

### Voice directions stack up too

"Rising, cracking on the last word, then dropping to a whisper" is three changes in one line. Video models that speak the lines play each change. Sighs and audible breaths are worse: written into a direction, they tend to come out loud and heavy.

## How to direct a natural performance

These are the rules we now give every writer in Movie. They work just as well when you write a prompt by hand.

### One feeling per person per shot

Pick the one thing each person feels in this shot. If the story needs a change, it happens at most once, on the story's turn: the moment a line lands or something is revealed. Never a chain of "then, then, then".

### Size it to the moment

Most moments are small. People show less than they feel, especially in front of others. Save shouting, sobbing, wide eyes, a big grin or laughing out loud for the story's peak, and give the peak to the person it belongs to. Across a short film that might happen once or twice.

### Write what the camera sees, not adjectives

Replace a stack of feelings with one or two physical cues the camera can film:

- Instead of "devastated, heartbroken", write: "she looks down and doesn't answer."
- Instead of "nervous and guilty", write: "he keeps his eyes on the screen; his jaw tightens."
- Instead of "overjoyed", write: "a small smile she can't hold back."

A cue gives the model something specific to draw, at a believable size.

### Let listeners listen

Someone who isn't speaking reacts quietly: they watch, they hold still, perhaps a small nod. That keeps attention on the speaker and makes the scene readable.

### Keep the character's temperament

A calm, wise or dignified character stays composed under pressure. That is often what makes them who they are. For a well-known character, follow how they are usually portrayed: their manner, their pace, how they speak to others.

### Direct the voice gently

Keep each person's usual pace, pitch and volume. Change it at most once in a line, and only as much as the moment needs. Raise it only at a real peak. Leave out sighs and breaths entirely.

## Before and after

Real directions from our test films, and how we would write them now:

| Before | After |
| --- | --- |
| Defensive anger that cracks, then resolve, then tired relief and quiet pride | Eyes on the screen, jaw tight. On her last line his shoulders drop and he nods once. |
| Wide-eyed curiosity that turns to pleading hope, then smug contentment | Watches the cup closely. When he gives in, one slow, satisfied blink. |
| Tearful pride (a character who only listens) | Listens, still, eyes wet. |
| Thunderous and booming, cracking into a snarl | Low and level; the last word hardens. |
| Grateful, luminous, urgent, face intent as she bows | Bows, eyes on him. Her voice quickens only on the warning. |

The "after" column is shorter, and that is the point. Each line gives the model one thing to show.

## A checklist before you render

- One feeling per person in each shot.
- At most one change, and only on the story turn.
- Physical cues instead of a list of adjectives.
- Listeners react quietly.
- Big displays only at the peak, for the person the peak is about.
- Voices at their usual level, with no sighs or breaths written in.
- A plain note for the model: "No exaggerated expressions: faces and voices stay the size real people show."

## How Movie handles it

[Movie](https://studio.oneinfer.ai/docs/movie-and-film) turns a story into shots and writes each shot's direction for you. Since November 2026, every step that writes those directions follows the rules above. A checker reads each shot's direction before anything is filmed and sends back any that overact; its corrections only ever make the acting smaller. Every shot with people in it also tells the video model, in plain words, to keep faces and voices the size real people show.

You still have the last word. Before you press **Start**, each shot is a card you can edit, so you can tone a moment down or let one land harder. See [Review the shots](https://studio.oneinfer.ai/docs/movie-and-film#review-the-shots).

![A movie's job screen: the story as shots, the cast, and a Start button with its price](https://studio.oneinfer.ai/docs/images/movie-plan.webp "A story ready to start, demo data")

Faces matter as much as acting: each actor in [Cast & World](https://studio.oneinfer.ai/docs/cast-and-world) keeps the same face and voice in every shot, so a quiet performance reads as the same person throughout.

## Questions

### Why does my AI character smile all the time?

Usually because the direction asks for a happy feeling without saying how much, or asks for several feelings that all read as a smile. Name one small cue instead, such as "a brief smile that fades", and remove words that ask for more.

### Should I just write "subtle" in the prompt?

It helps a little, but a specific cue works better. "A small smile she tries to hide" gives the model a size and a shape. "Subtle joy" still asks for joy.

### Does the video model matter?

Some models read directions more literally than others, but every model we have used plays a stack of adjectives big. The direction is the part you control, and it makes the biggest difference.

### How do I make a big emotional scene still feel real?

Build to it. Keep the shots before it small, give the peak to one person, and let everyone else react quietly. A big moment lands hardest after quiet ones.
