Multi-Character AI Video Prompts: Direct Two-Person Scenes
One character in an AI video clip is easy now. Two is where it falls apart. The wrong person says the line. Both characters do the action you gave to one. The man in the blue shirt ends up in the woman’s red jacket halfway through. Or the two faces slowly blend into a pair of cousins.
These are direction problems, not model problems. A film director blocking a two-person scene tells each actor where to stand, what to do and when to speak. Your prompt has to do the same job, in writing, for a model that takes every word literally.
This guide covers how to write that direction: naming characters so the model can tell them apart, placing them in the frame, assigning actions and dialogue to the right person on each model, and keeping the blocking stable when you cut between shots. If your problem is that a single character looks different from clip to clip, that is a separate topic, covered in the consistent characters guide. This one is about getting two or more people to behave correctly inside the same scene.
Rule 1: Give Every Character a Label the Model Can See
The model cannot see names. “Maya” and “Jake” mean nothing unless they are tied to something visible. Give each character a label built from a visual difference, and use that exact label every time.
Weak:
Two friends argue in a kitchen. She tells him he forgot again.
Strong:
The woman in the yellow raincoat and the tall man in a grey hoodie stand in a small kitchen.
Rules for labels:
- Contrast on at least two axes. Clothing color plus one of height, age, hair or build. “Woman in yellow, man in grey” is good. “Woman in a red shirt, woman in a pink shirt” invites a merge.
- Reuse the label word for word. If you call him “the tall man in a grey hoodie” in sentence one, do not switch to “her brother” in sentence three. The model may add a third person.
- Avoid pronouns when two characters share one. Two women in a scene means “she” is ambiguous every time. Repeat the label instead. It reads clunky to humans and clearly to models.
- Keep the count low. Two or three named characters is the practical limit for a single clip. Anyone else becomes background (“a few customers blurred in the background”) with no actions or lines of their own.
Rule 2: Block the Scene in Frame Terms
“They sit together” gives the model free rein on placement, and it will reshuffle them from take to take. Put each character somewhere specific, using the frame as your map:
- Left and right: “The woman in yellow stands on the left side of the frame, the man in grey on the right.”
- Near and far: “The grandmother sits in the foreground, the child plays on the rug in the background.”
- Facing: “They face each other across the table, both in profile to the camera.”
- Distance: “An arm’s length apart.” Leaving space between faces also reduces the chance of the model blending them.
Google’s Veo documentation lists “two-shot” as a composition term, and it is a useful shorthand for a frame that holds two people. Combine it with left and right placement: “Medium two-shot, the barber stands on the left behind the chair, the customer sits on the right facing a mirror.”
Rule 3: One Action Per Character Per Beat
The most common multi-character failure is shared action: you write “the dog barks and the man jumps back,” and you get a man who barks, or a dog that jumps. Separate the actions by character and by time.
- One verb per character per sentence. “The dog barks. The man in the blue shirt jumps back.” Two sentences, two subjects.
- Sequence with time words. “First… then… as soon as…” gives the model an order.
- Give the non-acting character a reaction. If you do not say what the second character does while the first acts, the model may copy the first character’s action onto them. “The woman stays seated, watching” is a complete instruction.
For clips with several beats, Google documents a timestamp format for Veo 3.1 ([00:00-00:02] ...), which is a clean way to say who does what and when. The storytelling prompts guide has more on sequencing beats across a longer piece.
Rule 4: Attribute Dialogue the Way Each Model Expects
Wrong-speaker lip sync is the most visible multi-character bug. Each model documents its own way of assigning lines.
Veo 3.1. Put speech in quotes and attach it to the character’s label with a verb. Google’s own example in the Gemini API docs is a good template: “A close up of two people staring at a cryptic drawing on a wall, torchlight flickering. A man murmurs, ‘This must be it. That’s the secret code.’ The woman looks at him and whispering excitedly, ‘What did you find?’” Notice the turn: the second speaker looks at the first before she speaks.
Kling 3.0. Kling’s VIDEO 3.0 guide says that if you clearly specify dialogue for each character, the model matches each character with their lines. Its format is a script line with a tone note in parentheses:
Mom (softly, in a surprised tone): Wow, I didn't expect this plot at all.
Dad (in a low voice, agreeing, in a calm tone): Yeah, it's totally unexpected.
With Kling 3.0 Omni, you can go further and create saved characters, called elements, then reference them with @Name in the prompt. Kling’s Omni guide shows voice binding too: attach a short speech sample to an element and its lines use that voice. Its guide lists up to 7 images or elements per generation without a video input, and up to 4 with one.
Seedance 2.0. Put spoken words in double quotes, and give each line to a labeled or referenced character (for example the person from @Image1). Keep lines short; long speeches tend to drift out of sync.
Rules that work everywhere:
- One speaker at a time. Write the lines in order with an action between them. Overlapping speech is rarely clean.
- Tell the listener to be quiet. “The man in grey listens silently, mouth closed” cuts down on both characters mouthing the same words.
- Keep lines short. Veo clips run up to 8 seconds, so two short lines is a full scene.
If one of the two characters is on camera talking and the other is only heard, the talking-head prompts guide covers the single-speaker framing side.
Rule 5: Keep the Blocking Stable Across Shots
Once a scene is longer than one clip, the second shot has to agree with the first about where everyone is.
- Respect the 180-degree rule. If the woman in yellow is on the left in the wide shot, she stays on the left in the over-the-shoulder and the close-ups. Write it into every prompt: “the woman in yellow, on the left side of the frame.”
- Repeat the full labels in every shot prompt. Each generation is fresh. The model does not remember that “the man” meant the tall one in grey.
- Use reference inputs for the cast. Veo 3.1’s ingredients-to-video takes up to three reference images, which fits two characters plus the setting. Google’s own multi-shot dialogue workflow does exactly this: one image each for a detective, a woman and the office, reused in every shot prompt that begins “Using the provided images for the detective, the woman, and the office setting…” Seedance 2.0 takes up to 9 reference images, and Kling 3.0 Omni has its element library.
- Use Kling’s multi-shot mode for shot-reverse-shot. Kling 3.0 can plan several shots inside one clip of up to 15 seconds, so a two-person exchange can be generated as one piece instead of stitched from separate generations.
For keeping each person’s face and outfit identical from clip to clip, go back to the consistent characters guide: locked descriptions, reference images and frame chaining all apply here, one character at a time.
8 Copy-Ready Multi-Character Prompts
1. Doorbell camera: package standoff
Doorbell camera footage, fisheye wide angle, daytime front porch. A delivery driver in a brown uniform stands on the left side of the frame holding a large box. A small white dog on the right side of the frame blocks the doormat, barking and refusing to move. The driver takes one step forward; the dog stands its ground and barks louder. The driver sets the box down slowly and backs away with both hands up. Slight compression artifacts, include barking and the driver laughing nervously.
More in this format on the doorbell camera prompts page.
2. Kitchen argument, two lines (Veo style)
Medium two-shot in a small apartment kitchen at night, warm overhead light. The woman in a yellow raincoat stands on the left side of the frame, the tall man in a grey hoodie leans against the fridge on the right. The woman holds up an empty milk carton and says, "You finished it. Again." The man in grey looks at the carton, then at her, and says, "Technically, there was a sip left." The woman stares at him silently. Sound of the fridge humming.
3. Street interview
Handheld phone video on a busy city sidewalk, afternoon sun. An interviewer in a black cap holds a microphone on the left side of the frame and turns it toward an older man in a tweed jacket on the right. The interviewer asks, "What's the best advice you ever got?" The older man thinks for a second, then smiles and says, "Never argue with a cat." The interviewer laughs; passers-by blur behind them. Street noise and wind on the mic.
The street interview prompts page has more setups like this.
4. Grandma and grandson baking
Medium two-shot in a sunny farmhouse kitchen. A grandmother with short white hair and a floral apron stands on the left, a boy of about seven in a blue t-shirt kneels on a chair on the right. First, the grandmother cracks an egg neatly into a bowl. Then the boy tries the same and crushes the whole egg in his hand. The grandmother pauses, then bursts out laughing while the boy stares at his dripping hand. Soft window light, sound of laughter and a spoon clinking.
5. Cat vs dog, one toy
Handheld phone video of a living room rug, low angle. A fluffy orange cat on the left side of the frame and a black labrador puppy on the right both stare at a single red ball between them. The puppy slowly stretches one paw toward the ball; the cat swats it once, quickly. The puppy freezes, then lies down flat. The cat pulls the ball toward itself without breaking eye contact. Natural daylight, include a soft thud and a small whine.
6. Three friends in a car (dashcam)
Dashcam footage from the windshield facing inward, night, city lights passing outside. The driver, a woman with a shaved head, is on the left; a man with round glasses sits in the passenger seat on the right; a third friend in a red beanie leans forward from the back seat, centered between them. The friend in the red beanie says, "We're lost, aren't we?" The driver and the man with glasses look at each other silently, then both burst out laughing. Grainy low-light dashcam look, road noise.
Three characters works here because the car fixes their positions for you. The dashcam prompts page has more.
7. Kling-style dialogue with tone notes
Medium two-shot, a quiet park bench in autumn, golden afternoon light. An old man in a flat cap sits on the left, a young woman with a guitar case sits on the right, facing slightly toward him.
Old man (gently, curious): You play every day here?
Young woman (shy, smiling): Only when it's not raining.
The old man nods slowly and looks back at the falling leaves. Soft wind and distant birds.
8. Detective and client, two-shot to close-up
Using the provided images for the detective, the woman and the office setting: medium two-shot, the detective sits behind his desk on the right side of the frame, the woman in a green coat stands in the doorway on the left. The detective looks up from his papers and says in a weary voice, "Let me guess. Someone's missing." The woman stays in the doorway, silent, gripping her handbag. Low-key lamp light, film noir mood, rain against the window.
Prompt 8 is built on Google’s documented ingredients-to-video workflow, so use it on Veo 3.1 with three reference images. On other models, drop the first sentence and keep the rest.
Troubleshooting Multi-Character Clips
| Problem | Likely cause | Fix |
|---|---|---|
| Wrong character speaks | Line not tied to a label, or pronoun ambiguity | Label plus speech verb before every line; add “the other listens silently” |
| Both characters do the action | Action written with a shared subject | One subject per sentence; give the second character a reaction |
| Outfits or faces swap | Labels too similar | Contrast clothing color plus one physical trait |
| Characters change sides between shots | No frame position in the second prompt | Repeat “on the left side of the frame” in every shot |
| An extra person appears | Label changed mid-prompt (“her brother”) | Reuse the exact label, word for word |
| Faces blend | Characters too close together | Add distance: “an arm’s length apart,” “across the table” |
The prompt mistakes guide covers the general errors that make these worse, like conflicting lighting and too many beats for the clip length.
Getting Multi-Character Variations Fast
The prompts that go viral in short-form are very often two or three characters in a tiny situation: a dog protecting its owner, a cat and a puppy fighting over a toy, a grandparent and a kid. LzyPrompt writes these with the blocking already done. Type an idea like “cat vs dog” and you get 10 variations, each with its own setting, who does what, action beats, sound cues and a caption. Here is a real output with three characters, each with a clear visual label and a clear job:
CCTV security camera footage in a dimly lit alleyway with faint street lighting, featuring a large brown dog and its owner, a young woman in a red jacket. A stranger in a black hoodie approaches, and the dog blocks the path, growling while the owner stands frozen, scared. The dog's fur stands on end as it positions itself firmly between the stranger and owner. Grainy security cam aesthetic, static surveillance, include low growls and footsteps sounds.
Copy the prompts into Veo, Kling, Seedance or any other model, or export all 10 to CSV.
FAQ
How many characters can an AI video model handle in one clip?
Two is reliable, three is workable if each has a clear label and position, and beyond that the extra people should be background with no individual actions or lines.
How do I make the right character say the right line?
Attach every line to a visual label with a speech verb (“The woman in yellow says, ’…’”), write lines in order, and tell the listener to stay silent. On Kling 3.0, use its script format, Name (tone): line.
Should I use names like “Maya” and “Jake” in prompts?
Only if the name is tied to a visual description or a reference, like a Kling Omni element. On their own, names do not tell the model who is who.
How do I keep two characters on the same sides across multiple shots?
State their frame positions in every shot prompt and keep them consistent with the first wide shot, the same way film crews follow the 180-degree rule. Reference images for each character help keep their looks stable too.
Direct Your Next Two-Person Scene
Multi-character clips work when every person has a visible label, a place in the frame, one action per beat and their own lines. Write it like a director’s blocking note and the model follows.
When you want 10 ready-blocked variations of your next idea, try LzyPrompt free for 7 days. It is $9 a month or $60 a year, and prompts are unlimited.
Bank K.
Founder, LzyPrompt
Builder of LzyPrompt. Creates AI video prompts to help content creators save time generating professional videos for YouTube Shorts and Facebook Reels.
@ifourth on XRelated Articles
AI Video Aspect Ratio Prompts for 4 Ratios
Copy-paste aspect ratio prompts for 16:9, 9:16, 1:1 and 21:9, plus when to use each, framing tips, model notes and common mistakes to avoid.
Read more →AI Video Camera Movement Prompts: 8 Moves
Copy-paste prompts for 8 camera moves: static, pan, tilt, dolly, tracking, crane, push-in and orbit. Plus speed words, prompt order and templates.
Read more →Ready to Try LzyPrompt?
Create professional AI video prompts in seconds. Start your free trial today.
Start Free Trial