Two-character AI video prompts: give each person a clear action.
Identify the two people in the opening view, attach each movement to the person who performs it, and say who handles any shared object. Read the prompt back as a small piece of choreography: you should be able to point to the actor, the action, and its result.
This guide is about two people acting within one shot. For carrying the same person across separate clips, use the character consistency guide. Clear wording communicates a plan; it cannot guarantee that a model will preserve identities, hands, or physical contact correctly.
1. Make the action readable before adding style
“Two people examine a toy and lift it” leaves an important decision open: does one person lift it, or do both? Replace that sentence with the intended behavior of each person. If the distinction matters to the story, make it visible and easy to check.
Runway's Gen-4 Video Prompting Guide recommends positional language or simple descriptive identifiers when subjects in an image need different movements. That is guidance for Gen-4, not a universal identity-lock feature. Choose labels that actually distinguish the people in your input.
The original planning exercise below imagines two adults at a repair bench. A repairer in a charcoal apron is on the left; a visitor in a teal jacket is on the right. A small wooden bird rests on the bench near the repairer. No reference image or generated video accompanies this example, and these drafts have not been tested on a video model.
On a small screen, scroll the table sideways.
| Element | Intended behavior | What to inspect |
|---|---|---|
| Repairer, initially screen-left | Lifts the wooden bird just above the bench and holds it. | The repairer performs the lift, with the bird remaining visibly supported. |
| Visitor, initially screen-right | Watches the bird, keeping both hands near their own body. | The visitor does not become the person holding the bird. |
| Wooden bird | Moves from the bench into the repairer's hands. | One recognizable bird remains through the lift. |
| View | Holds a medium view of both people and the bench. | The action remains visible rather than leaving the frame. |
This is a review worksheet, not a format the model must parse. Names such as “repairer” and “visitor” are plain descriptions. If both people wear similar clothes or change screen positions, those descriptions may need a clearer visible cue. Screen-left is a location in the view; it is different from a person's own left hand.
2. Check whether the people are already in the input
With a first-frame image, inspect that image before writing movement. Are both people present? Can you tell them apart? Is the bird in a place where the proposed lift makes sense? If the opening image already has the visitor holding it, the motion brief above starts from the wrong situation.
Runway's current image-to-video guide explains that the input image supplies the scene's composition and appearance, while the text mainly directs motion. It also notes that interactions between elements can benefit from visual descriptions. Use enough detail to resolve the interaction without treating every visible feature as something to redesign.
For text-to-video, establish the people, their arrangement, and the object in the draft because there is no starting image supplying them. For a service with separate character-reference inputs, check what those controls accept. Two files in an upload area are not necessarily two independently controlled characters. Use the service's actual input rules; filenames in your notes are not automatic reference syntax.
If you need help describing what a starting image really shows, use the reference-to-words guide. Keep observations separate from details you want to add.
3. Draft the same scene for two different inputs
Motion draft for an already approved starting image
The apron-wearing repairer on the left lifts the wooden bird from the bench and holds it just above the surface. The teal-jacketed visitor on the right watches the bird, with both hands resting near their own body. The camera holds a medium view of both people and the bench. The shot ends with the repairer holding the bird.
This assumes the imagined starting frame already has the arrangement described above. If your real image differs, adapt the identifiers and starting state. Do not paste the draft over a contradictory input and assume that the role words will repair it.
Scene draft when starting with text
A medium view of two adults at a wooden repair bench in a softly lit workshop. An adult repairer in a charcoal apron stands on the left. An adult visitor in a teal jacket stands on the right. A small wooden bird rests on the bench beside the repairer. The repairer lifts the bird and holds it just above the surface. The visitor watches with both hands near their own body. The camera holds its position.
Compare the drafts by their jobs, not their length. The first asks for movement in an established image; the second also describes the scene to create. Neither includes dialogue or a camera move that would add another action to review. You can plan a different viewpoint with the camera movement guide.
Read it backward as a role check
Ask “Who finishes holding the bird?” before asking whether the take looks cinematic. The answer should be the repairer. If your intended answer is the visitor, change the choreography explicitly: a handoff introduces a different interaction with its own contact and timing to inspect. Simply adding “they exchange it naturally” hides those decisions again.
4. Review the interaction in order
- Opening: identify both people and the bird before the lift.
- Approach: check whose hands move toward the object.
- Lift: look for support, changes of shape, duplicate objects, or an unexpected person taking over.
- Hold: confirm who has the bird and whether the other person still reads as the same visitor.
Watch the full completed clip, then scrub the moment that looks wrong. Record what you see, such as “the visitor reaches in during the lift,” instead of guessing at a model's internal cause. A face change, a wrong actor, and an object that loses its shape are different observations and may call for different next attempts.
Save the model and input mode, approved source files, exact prompt, settings, output identifier, and the point where the intended action diverges. If you revise one instruction, keep the other choices recorded so the comparison remains useful. Do not claim that one successful take proves the wording will work every time.
5. Decide what the scene actually needs to show
If the story only needs the visitor to see the repaired bird, try a version where it is already in the repairer's hands and the visitor looks toward it. That changes the planned action and avoids asking for the lift. Review the new starting image and result rather than calling it a fix for the original take.
If precise contact is essential and repeated attempts miss it, plan separate views or use an editing or compositing workflow you can control. A close view of the object followed by the visitor's reaction can communicate a different version of the scene; it does not demonstrate a successful continuous handoff. Use a storyboard card to record that choice before making more clips.
You can edit, save locally, and export your motion draft in Shotmelo's image-to-video prompt planner. Keep the action worksheet with your scene notes. The current preview uses manual templates; it does not analyze people in an image, render video, or assign persistent character identities. Run the final draft in your chosen video service and judge the actual result.