How to Put Multiple Characters in One AI Video Scene (and Keep Them Straight)

To put multiple characters in one AI video scene and keep them straight, give every character a reference sheet, stage the scene before you generate it (who is where, who is screen-left and who is screen-right), keep the camera on one side of the 180-degree line, cover a conversation with an establishing wide shot followed by singles and shot/reverse shot, and assign every line to its speaker so lip sync lands on the right face. Paintbrush does the staging and coverage automatically: its AI cut director stages each chapter, draws a camera map, and cuts the recorded conversation into shots that each list who is in frame.

Why do AI videos with two or more characters break?

One character in one clip is the easy case. Add a second person and a cut, and four failures show up:

• Identity blending. The model mixes features between people. She gets his jacket, or both end up with the same face. It happens most when characters are described only in text and look alike. See why AI video loses character consistency.
• Swapped positions between cuts. In shot one, Mara sits on the left and Theo on the right. In shot two, a new angle, they have swapped. Each clip is generated on its own, so the model has no memory of where people were. The conversation suddenly looks mirrored.
• The wrong mouth moves. Two faces on screen, one line of audio. If the tool does not know who is speaking, the listener may mouth the words.
• Extras appear and disappear. The waitress behind the counter is in the wide, gone in the next angle, then back. Or a stranger nobody wrote shows up in the next booth. Models fill empty space with people unless told exactly who is there.

How do you stage a scene before generating it?

Staging means deciding where everyone is before any camera exists. Write down where each character is relative to the furniture and each other, who arrives or leaves and from which direction, anyone present but noticed later, and the time of day and light.

Then make a camera map. For each conversation it answers three questions:

1. What is the line of action? The imaginary line between the two people talking. In a diner booth it runs across the table.
2. Which side does the camera keep? Pick one side of that line and stay there for the whole conversation. This is the 180-degree rule. If the camera never crosses the line, the person on the left stays on the left.
3. Who is screen-left and who is screen-right? Mara screen-left, looking right. Theo screen-right, looking left. In their singles, each looks toward the side where the other sits.

Last, name the two to four views the scene uses, such as "toward the window" or "toward the counter". Named views stop each shot from inventing a new room behind the characters. For keeping the location itself stable, see how to keep backgrounds consistent in AI video.

Which shots should you use for a conversation?

Use the coverage film and TV use. Each shot has one job:

• Establishing shot. A wide showing the space and everyone in it. It gives the audience the geography once.
• Two-shot. Both speakers in one frame. Best for physical moments between them: a reach, a touch, someone pulling away.
• Singles. One person, medium or close-up. Most dialogue lives here, because the speaker's face and mouth are clear.
• Shot/reverse shot. Alternating singles of each speaker, filmed from the same side of the line.
• Reaction shots. The listener's face as a line lands. Often the most important shot in the scene.

Start wide and move closer as the emotion rises. Keep dialogue shots to 2 to 6 seconds. Weighty moments, like an entrance, a touch or the turning line, get their own shot. A line and its reaction can share one.

How does Paintbrush keep multiple characters straight?

Paintbrush builds each scene as a chapter: one continuous stretch of story in one setting, written as a script and performed as one piece of audio.

References carry identity, not text

Every character gets a character sheet (front, side profile, back and a face close-up) for 36 credits. In a shot you reference characters by @mention, like @Mara and @Theo. Paintbrush sends each mentioned character's images to the video model, face first, then front, side and back. Text descriptions of their looks are not injected into the prompt. More in the character consistency deep dive.

The cut director stages and covers the chapter

Once a chapter's audio is recorded, an AI cut director cuts it into shots, for free. It stages the chapter first: everyone in the space, including people noticed later, positions, arrivals, departures and light. Then it writes the camera map: line of action, which side the camera keeps, who is screen-left and screen-right, and the named views. Then it cuts coverage: an establishing shot, closer shots as emotion rises, shot/reverse shot, separate shots for weighty moments, and a clear face and mouth for the on-screen speaker. Each shot lists who is in frame, so a wide shows everyone present. Characters look at each other, not into the lens, unless the shot calls for direct address.

New angles start from the previous shot's real ending

Shots generate in order, each staged from the previous shot's actual ending. A cut to a new angle in the same space is first drawn as a still keyframe from the earlier shot's real last frame, then animated, which keeps posture, positions, clothes, room and light. The keyframe adds 8 credits. Before generating, a vision model checks which side each character actually ended up on and corrects the shot description, so the axis is not jumped.

Lines land on the right face

Every script line has a speaker, so the system knows whose mouth should move. As of October 2026, every project, in every style, generates video with Seedance 2.5; Paintbrush's Seedance setup accepts photoreal character references, so realistic casts work too:

• On-camera lines are performed by the video model itself. For each shot, Seedance hears that slice of the chapter's recording and performs the lines on the speaker's face, voice and mouth together. There is no separate lip-sync step or charge, however many people speak in the clip.
• Narration and offscreen lines move no mouth, so the listener on screen stays quiet while the voice plays.

Video models also invent stray voices. On clips nobody speaks in on camera (silent, or narration and offscreen lines only), Paintbrush removes the generated voices, keeps the ambience and lays the approved narration back exactly. Clips with on-camera dialogue keep the video model's own soundtrack.

Clearly different voices

When a line plays over a reaction shot, the voice tells the audience who spoke. Voices come from the ElevenLabs Voice Library, and the creation agent gives characters who share a scene clearly different voices. The chapter is performed as one take with ElevenLabs Eleven v4 Text to Dialogue, so each voice hears the others. See how to make AI characters talk.

What are the best practices for multi-character scenes?

• Two or three speaking characters per shot. More can sit in the wide. Cover everyone else with singles and reactions.
• Distinct silhouettes and colors. Different heights, builds and hair shapes, and one signature color each. A red coat and a grey hoodie are easy to tell apart, for viewers and for the model.
• Staging written with @mentions. "@Mara sits on the window side of the booth. @Theo sits across from her, back to the door."
• Offscreen lines marked. They move no mouth, so a reaction shot can carry a line from someone out of frame without the wrong person mouthing it.
• Silent beats for physical moments. A reach or a look away written as a silent beat can get its own shot.
• Costume changes in outfits, not shot descriptions. See how to change a character's outfit in AI video.

What does a multi-character scene look like in practice?

A short chapter with three characters, one offscreen line and two silent beats.

Setting: @Night Diner. Staging: @Mara sits on the window side of a booth, facing the door. @Theo slides in across from her. @Ines works behind the counter and walks over late in the scene. Late night, warm overhead light.

1. Silent beat, 3 seconds: Theo slides into the booth and sets an envelope of cash on the table.
2. Theo: [nervous, too fast] I sold the car.
3. Mara: [flat] You sold the car.
4. Theo: It was that or the rent. [pause] I didn't want to ask you again.
5. Silent beat, 2 seconds: Theo reaches for her hand. She pulls it back.
6. Ines (offscreen): [calling from the counter] Refills over there?
7. Mara: [quietly, eyes on Theo] We're fine.
8. Ines: [dry, warm] Honey, nobody in this booth is fine.

Camera map: the line runs across the table and the camera keeps the counter side. Mara is screen-left, Theo screen-right, and Ines enters from screen-right. The shot list:

Shot 1 shows Ines early, so her entrance makes sense. Shots 2 to 4 are shot/reverse shot from one side of the line, so Theo always looks left and Mara always looks right. The refused touch gets its own two-shot. Shot 6 pairs Ines's offscreen line with Mara's reaction, and only Mara's mouth moves. You can split, merge or re-time any shot.

What does a multi-character conversation cost?

As of October 2026, prices are the same in every style. Video costs 30.3 credits per second. A 5-second shot with on-camera dialogue costs 151.5 credits, lip sync included. A 5-second silent or narrated shot costs 164, because it also pays 12.5 for voice cleanup. Because the speech is performed inside the video, a two-shot where both people talk costs the same as a single. Clips run 4 to 30 whole seconds, so a 2-second reaction still pays for a 4-second clip (121.2 credits with dialogue, 133.7 without). Each new-angle keyframe adds 8. Speech for a one-minute conversation is under 10 credits, and cutting is free. After a one-time $3 intro pack of 100 credits, packs start at $12 for 500. Full math in what AI video costs per minute.

Can you do this in other AI video tools?

Yes. In a general-purpose clip generator you write the staging into each prompt, check sides by eye and handle lip sync per clip. Those tools are best for single striking shots you control by hand. A tool built around saved characters and scripts is best for recurring people who talk to each other. See the best AI video generators for consistent characters.

FAQ

Can AI video show two characters talking to each other?

Yes. Give both characters references and voices, write the conversation with a speaker on every line, and cover it with an establishing shot and shot/reverse shot. Paintbrush performs the conversation as one take, cuts it into shots, and the video model performs each on-camera line on the right face.

How many characters can be in one AI video scene?

As many as the wide shots need. Keep two or three speaking characters per shot, each with a distinct design and voice.

How do I stop AI characters from switching sides between shots?

Keep the camera on one side of the line of action and note who is screen-left and screen-right. Paintbrush also draws each new angle from the previous shot's real last frame and checks screen direction before generating.

Do I have to write the shot list myself?

No. The cut director stages, maps and cuts the chapter for free. Edit any cut, or ask the creation agent, for example "make the argument a shot and reverse shot". For the lines themselves, see how to write dialogue for AI video.