Case study
Why an AI image adds an extra person to your two-person scene
A client's video team wanted a two-person conversation ad. The AI image tool kept drawing the person each speaker was looking at, so the cut would show four people. Describing the framing, not who must not be there, fixed it.
Key takeaways
- An AI image draws what your description says, so a phrase about someone off the frame can put a stranger in the picture.
- To keep one person alone, describe the framing: a three-quarter turn, open space in front of the face, and edges that show the place rather than another person.
- A two-person conversation needs each speaker looking at the other, not into the lens, with the two directions matching across the cut.
- Long AI video scenes should be ordered a few clips at a time, with the next asked as each finishes, and a failed clip should be a decision for a person.
A client's in-house video team can now make a two-person conversation ad inside the studio it uses every day, with each speaker drawn alone and looking at the other. The conversation feature was built within a day of the team asking, and the fixes for the pictures followed the same day. The problem it solved is the one you meet whenever an AI image adds an extra person to a scene you meant for two: the tool draws the words in your description, not your intention.
A conversation opening only works if the two people look real
A conversation is one of the strongest ways to open a video ad, and it fails at once if the two people do not look as if they belong in the same place. The first seconds decide whether anyone keeps watching. It only works when the same people are in the same place and each is looking at the other.
A cut that suddenly shows strangers, or a scene that falls back to stock footage halfway through, breaks it. The client's team had made its conversation by hand in a separate AI tool. The speaker did not sound like the presenter of the main ad, and the two halves did not match.
Why an AI image adds an extra person to your two-person scene
An AI image tool treats every word in your description as something to draw, so asking for a person looking at someone just off the edge of the frame gives it a someone to draw. A stranger then appears over a shoulder, or beside the speaker.
Adding a line saying nobody else is in the picture does not help. The word still puts another person into the description. The same principle applies to video when you ask for more clips at once than the tool can make: the late ones can wait past the time allowed and fail.
What the first live pictures showed
The first pictures, made on the live studio, added the very person each speaker was looking at. A stranger appeared over a shoulder, a woman stood beside the speaker, and a second man stood beside the man he talked to. The cut between the two would have shown four people instead of two.
We added a line saying nobody else was in the picture. The next pictures still showed a person over the shoulder.
A second fault appeared when the team made a longer conversation. Every line for the second speaker was sent at the same moment. The video tool made the few clips it could at once, and the last lines waited past the time allowed for a clip and failed. The finished ad showed stand-in stock clips where those lines should have been.
What we changed
We stopped describing who must not be in the picture and described the framing that leaves room for only one person. Each person is now described as turned a little to one side in a three-quarter view, looking that way, placed off the centre with open space in front of the face, and with the edges of the frame showing the place, not people. The person asking looks right; the person answering looks left. Drawn that way, each came out alone, and the two cut together as a conversation.
The studio now reads a conversation script as a screenwriter writes it: each turn opens with who speaks, and stage directions sit in brackets so nobody reads out a name or a direction. Each role is given to a presenter, and a team screen lists who plays whom; the ad cannot be approved while a role has nobody. Each person is drawn in the conversation's place, or in a place of their own when the script says so, and production waits until every picture is ready.
The second speaker's lines are asked for a few at a time, and the next is asked as each finishes or fails, so none waits past its limit. A line that still fails is left for a person to decide, not quietly replaced with stock footage. Pictures made under the old description are drawn again rather than reused.
What the business has now
The client's team now makes conversation ads in the studio it already uses every day. Since the framing change, every speaker comes out on their own and faces the right way. Longer conversations are now filmed one line at a time, with no stock stand-ins.
The fix for the late lines came a few days later, the day the team made a longer conversation.
What you can check yourself
- Look at each cut of a two-person scene on its own. Count the people and see where each is looking: a face turned to the camera, or a third figure at the edge, is a sign.
- When an AI picture keeps adding something, read your description for any word that suggests it, even inside a no. Rewrite the framing rather than adding another prohibition.
- Watch a finished AI video to the end. If a speaker is replaced by stock footage, a clip failed without anyone deciding it should.
- Ask your video tool how many clips it can make at once and how long one may take before it gives up. A scene sent in one batch can run past that limit.
What to do next
Start with the shot that looks wrong, not with another line telling the tool what to avoid. If two people should look at each other, describe each shot with the turn of the head and the space in front of the face. Keep each speaker's lines with that speaker, and pace a longer scene in small batches.
If your team makes video ads with AI presenters and wants a two-person conversation that holds together, this is the kind of work our engineers do for clients.
Questions people ask
Why does my AI image keep adding a person I did not ask for?
The tool is not reading your intention; it is drawing your words. If the description mentions someone just off the frame, the tool has a reason to draw that someone. Saying nobody else is there can still put another person into the description. Describe the framing of the one person instead.
Does telling an AI image tool not to draw someone work?
It is not reliable. A prohibition still names the thing, and the image tool treats the words as material to draw. In the work here, a line saying nobody else was in the picture still produced a person over the shoulder. The change that worked was to describe the subject's turn, position and the space around the face.
How do I make two AI presenters look like they are talking to each other?
Draw each presenter alone, looking towards the other, and cut the two shots together. Have the person asking look one way and the person answering look the opposite way, in the same place and with a consistent look. Neither should look into the lens if the scene is meant to be a conversation between them. Keep each line with its speaker so the voices and faces stay matched.
Why did some lines in my AI video come out as stock footage?
A long scene was probably sent in one go. The video service can make only a few clips at once, so the late lines can wait past the time allowed and fail. If a failed line is replaced automatically, the finished ad can show stock footage where a speaker should be. Ask for lines a few at a time and have a person decide what happens to a line that fails.
How many AI video clips can I ask for at once?
The limit belongs to the video service you use, so ask whoever looks after it rather than guessing. Send a long scene in smaller batches and start the next clip as each one finishes. That keeps the late lines from waiting past the limit.