OmniHuman 2.0 multi-person prompts should identify where each person stands, who speaks first, and what the listener does during each turn. BytePlus lists multi-person scenes for its film-grade tier, but doesn’t publish a promise that a text prompt can assign every syllable to the right face. Start with a single two-person shot and short, clearly separated lines.<\/p>\n
BytePlus says OmniHuman 2.0 works from one image plus audio and text, and lists multi-person performances among its capabilities. That means the reference image does a lot of casting work. If you need a host on the left and a guest on the right, put both in the same source image with their faces visible. Don’t expect a sentence about "the guest" to create a reliable identity when the photo gives the model no clear guest to follow.<\/p>\n
Leave room between faces. A chest-up two-shot with a little space between shoulders makes it easier to judge whose mouth moves than a crowded group photo; that’s a framing recommendation, not a published BytePlus requirement. Start with a blank wall. If you’re trying to assess a dialogue, an animated crowd or a complicated camera move gives you more things to blame when a turn looks wrong.<\/p>\n
Make the handoff audible. Cut overlapping lines in the first pass, and leave a small beat before the reply. This is an editing tactic for a human-readable conversation, not a claim about supported audio formats, the ability to separate voices automatically, or any guarantee that a given line will appear on the face you specified. BytePlus’s public product page does not spell out a multi-speaker track layout or say that separate voices will always map to separate people. Check the current API documentation for input constraints before you build an automated workflow. The API access guide<\/a> covers the access route without guessing at console menu names.<\/p>\n Treat these as starting templates. They aren’t tested outputs. Replace each role with a description that matches someone in the image. Prompts can steer action, but they can’t repair an unclear photo or a badly cut recording.<\/p>\n Interview with restrained reactions<\/strong><\/p>\n Fixed medium two-shot in a quiet studio. The person on the left asks the first question and then listens with a small nod. The person on the right answers while facing slightly toward the interviewer. Keep both people visible and preserve their positions throughout.<\/p>\n<\/blockquote>\n The listener has one small job. Watch the handoff, not the background. If the answer appears on the wrong mouth, don’t add five more adjectives; simplify the audio and cut to separate shots.<\/p>\n Back-and-forth product discussion<\/strong><\/p>\n Two colleagues stand beside an unmarked product on a plain table. The person on the right speaks first and gestures toward the table. The person on the left waits, then responds with one brief hand gesture. Keep the camera still and the product in the same place.<\/p>\n<\/blockquote>\n This separates the two turns with a change of physical focus. It doesn’t ask the model to invent packaging, a readable label or a new person. For an ad, avoid making the on-screen person appear to endorse a product without their permission.<\/p>\n Conversation with a quiet listener<\/strong><\/p>\n Wide, steady two-shot. The person on the left speaks for most of the clip. The person on the right remains silent, looks toward the speaker, and gives one subtle nod after a pause. No camera move and no new people entering the scene.<\/p>\n<\/blockquote>\n Use it for context, not a duet. The stillness makes an accidental speaking face easier to catch. If the listener begins mouthing the narration, a single-person shot edited against a separate reaction shot may be the cleaner production choice.<\/p>\n For broader motion phrasing, see the camera and gesture prompt examples<\/a>. Keep that extra motion out of the first multi-person attempt.<\/p>\n A requested turn isn’t a verified turn. BytePlus describes the film-grade tier as supporting multi-person scenes and precise gesture control, but its product page does not describe deterministic speaker-to-face assignment for dialogue. A prompt such as "left person speaks first" is a direction to try, not a technical guarantee.<\/p>\n Break a script into chunks with clean pauses before you spend money on a long take. Write down the first audible word for each speaker. Keep it simple. When reviewing the result, mute it once and watch the faces, then play it with sound and check each entrance. That review method catches an easy-to-miss problem in an otherwise plausible two-person shot: both faces start moving at the moment only one voice speaks, and the error becomes obvious as soon as you watch without sound. It isn’t a benchmark, and no video has been generated for the examples here.<\/p>\n If you must have exact speech ownership for a client approval, plan an edit-friendly route from the start: generate one speaking character at a time and assemble the exchange in your editor. You lose the continuous shared shot, but you gain control over where each voice appears. The lip-sync guide<\/a> is a better place to start when the mouth matters more than the shared scene.<\/p>\n BytePlus lists the film-grade tier at $0.07 per second and says its video output range is 30 to 300 seconds. A 60-second two-person scene is 60 s \u00d7 $0.07 = $4.20 at that published rate. Three separate 60-second generations would be 3 \u00d7 $4.20 = $12.60 before any other charges or changes in billing terms. The pricing guide<\/a> goes through the per-second math in more detail.<\/p>\n My vote: spend the first paid attempt on a plain, static conversation. A swooping camera plus two active hands is a poor way to learn whether your audio handoff works. BytePlus also lists an avatar tier at $0.12 per second for lip-sync-oriented avatar work, but the public pricing card doesn’t promise that choosing it will solve multi-person speaker mapping. Compare the tier descriptions against your shot before buying extra attempts.<\/p>\nPrompt patterns for a two-person exchange<\/h2>\n
\n
\n
\n
Keep speaker assignments honest<\/h2>\n
Budget for retries, not just runtime<\/h2>\n