For a two-person shot, give each person a recognizable description, assign a screen position, and state whose action happens first. Start with one simple action at a time, keep routes from crossing, and inspect the opening, movement, any occlusion, and the ending. A prompt expresses intent; it does not guarantee that identities, limbs, or actions stay locked.
What can go wrong in a two-person shot?
Identity swap means Person A’s stable traits appear on Person B. Action misassignment means the other person performs the requested action. A duplicate person can appear near an edge, reflection, or background. Body fusion occurs when arms or silhouettes merge. A post-occlusion change appears after one person passes behind another or an object. These symptoms need different checks; they are not all caused by a prompt being too short.
Build a two-person action plan
A and B are writing labels only, not character IDs or region locks. Screen-left and screen-right mean the viewer’s sides, not a person’s own left and right.
| Person label | Stable appearance | Starting position | One main action | What the other person does | Ending position |
|---|---|---|---|---|---|
| A | Blue jacket, dark trousers, short hair | Screen-left | Walks forward first | B waits with arms down | Screen-left foreground |
| B | Beige coat, light trousers, long hair | Screen-right | Walks after A | A has stopped on screen-left and remains still while B walks. | Screen-right foreground |
Use a clear two-person source with separated silhouettes, distinguishable clothing, simple starting poses, and permission to use the image. Avoid severe overlap, mirrors, cropped hands, and complicated poses. Two separate portraits do not guarantee a correct merge. For single-character context, see AI video character consistency; for general checks, see prompt troubleshooting and the shot-list template. Start from Seedance 2.5.
Prompt examples for two people
These are teaching examples to adapt and inspect, not tested success cases.
Two people walk through a wide corridor
Source condition: Wide corridor image with two fully visible people and a clear floor path.
Use the uploaded two-person corridor image as visual reference. Person A wears a blue jacket and dark trousers and starts on screen-left. Person B wears a beige coat and light trousers and starts on screen-right. Keep their clothing, faces, body proportions, and starting positions distinct. First, A walks forward along the left half for two steps while B waits still on the right. Then B walks forward along the right half for two steps. Both move in the same direction on parallel paths that never cross. Keep the camera fixed, preserve corridor geometry, and do not add a third person.
Observe A’s two steps before B’s and the separate routes. If it fails, remove B’s motion or shorten the route first; change one action assignment.
One person speaks, the other listens
Source condition: Medium two-person shot with both faces visible and no face occlusion.
Use the uploaded two-person indoor image as visual reference. Person A sits on screen-left in a green sweater. Person B sits on screen-right in a gray shirt. Keep their clothing, faces, and seats distinct. A turns slightly toward B and says, “Let’s take the next turn.” A then closes their mouth. Only after A finishes, B looks at A and gives one small nod while remaining silent. Keep the camera steady and avoid overlapping gestures. If audio is generated, the voice and mouth movement belong to A during the sentence.
Observe the speaking turn, mouth closure, and delayed nod. If it fails, remove the spoken line and test silent turn-taking first.
Foreground subject and passing background pedestrian
Source condition: Street or shop-window image with one foreground person and a simple distant pedestrian path.
Use the uploaded street-window image as visual reference. Person A is the only foreground subject, standing on screen-left in a red coat and facing the shop window. A stops, keeps both feet planted, and turns their head slightly toward the display. In the background, one different pedestrian in a dark coat takes two steps within the screen-right background and remains visually separated from A’s silhouette throughout. Do not duplicate A. Keep the camera fixed and add no extra people.
Observe A’s stationary action and the separate background route. If it fails, remove the background walk, establish A, then reintroduce one distant pedestrian with a shorter path.
Rewrite ambiguous actions
Ambiguous: “Two people talk, turn around, and walk away together.” A clearer version is: “Person A stands on screen-left and says one short sentence while Person B listens without speaking. After A closes their mouth, B turns toward the corridor. One beat later, A turns. A walks away along the left side; B follows on the right. Their paths remain parallel and silhouettes do not overlap.” The revision assigns an actor, order, position, and route to every action; it is still not a guarantee.
Troubleshooting table for multi-character shots
| Symptom | What to check | Smallest next adjustment |
|---|---|---|
| Identities exchange | Are descriptions tied to clear positions? | Repeat appearance and position; simplify action |
| Both do the same action | Is the waiting person’s state explicit? | Name the active person first; describe the other as still |
| Bodies fuse | Do contact or crossing routes occur? | Increase spacing, remove contact, or shorten shot |
| Appearance changes after occlusion | Is occlusion long or complex? | Shorten it or split into two angles |
| Extra person appears | Are background people, reflections, or cropped figures present? | Simplify background and request no additional people |
These are directions for investigation, not guaranteed single causes. Compare start, action, after-obstruction, and ending frames.
When to stop adding prompt text
If attempts keep swapping identities or fusing bodies, split into a single-person shot, reaction shot, or shot/reverse-shot pair, then combine clips in an external editor. A longer prompt cannot substitute for simpler interaction when exact contact or timing matters. Reference quantity does not equal a guaranteed number of controllable people.
Multi-character acceptance checklist
| Check point | Verify |
|---|---|
| Opening | Correct number, identity, screen position, starting pose |
| During action | Named person performs named action; other state is correct |
| During occlusion | No duplicate body, swap, or clothing change |
| Ending | Correct number, positions, silhouettes, contact boundaries |
Pause at the opening, action midpoint, first frame after obstruction, and ending. Do not judge only the first frame.
FAQ
Can a prompt guarantee that two people never exchange identities?
No. Clear labels and simpler routes help inspection but do not guarantee identity locking.
Should I provide a separate reference for each person?
Use the clearest available source. A well-composed two-person image shows the relationship directly; separate portraits do not guarantee a correct merge.
Why is a two-person shot harder than a single-person shot?
It must preserve two identities, silhouettes, spatial relationships, and action ownership at once. Occlusion adds more opportunities for visible changes.
What should I simplify after a failure?
Reduce simultaneous actions, remove crossing or contact, shorten occlusion, and let one person act while the other waits.
