Why does AI video add people or props I never wrote?

Updated 2026-10-27·3 min read·Consistency

Where the extra people come from

A video model starts from a picture and a prompt. If the picture has a wheelchair in the corner and the prompt doesn't mention it, the model animates it along with everything else. A person standing in the back gets the same treatment.

The reverse happens too. If the text names someone who isn't in the frame, the model may invent them, walk them in, or ignore the line.

Start frames pick up extras in a few common ways:

• The frame was taken from an earlier shot that had more people in it.
• The room still was drawn with furniture or props the scene doesn't need.
• A reference you attached brought its own background along.

What we caught

In our episode Marry My Broken Son, the checks under one start frame flagged a wheelchair and Maria. Both were in the picture, and the beat's text named neither of them.

That's the moment to stop. Rendering would have meant paying for a shot where the model had to decide what to do with a wheelchair and a bride nobody asked for.

The fix

1. Read the checks under the start frame. They compare the picture with your beat and list what's in one but not the other.
2. Decide which one is right. Is the frame right and the text incomplete, or is the text right and the frame carrying extras?
3. If the text is wrong, add what's there. Tag the person or prop in the beat so the model knows they belong.
4. If the frame is wrong, change the picture. Pick a different start frame, or use Shoot with AI's Change and say what to remove.
5. Fix it before you render. A still is cheap to fix. A video take with the wrong people is a full re-render.

Write beats that match the frame

Each beat is one camera setup. Say what happens, tag the people, rooms and props, and put the line in quotes after the speaker. A few habits help:

• Tag everyone who is visible, including people who don't speak. If they're in the frame, they belong in the beat.
• Tag outfits by their own tag, so the right costume renders in that beat.
• Keep one camera per beat. When a beat says “cut to,” split it. A warning flags a beat trying to be two setups, and Split gives each its own beat.

Why the checks matter more than the prompt

You can write a perfect prompt and still get extras, because the picture outranks the words. The model sees the frame first. Checking the frame against the text is the cheapest place to catch this, and it's before anything is spent on video.

For keeping the same faces across shots, see AI character consistency. For rooms that drift, see AI location consistency.

See the checks in action in the full guide and the Short Same AI characters, every scene.