OmniHuman2
OmniHuman 2 prompts work best when they describe a shot someone could actually direct: who moves, what the camera does, and when the movement should settle. Start with one visible action and one camera instruction, then let the uploaded audio carry the speech. Treat the examples below as editable shot plans, not tested recipes or guaranteed outputs.
Give the camera one job
BytePlus describes OmniHuman 2.0 as taking a single image with audio and optional text, and says prompts can control camera movement. It mentions gestures, too. Its linked OmniHuman 2.0 prompt guide says you can describe several actions as well as camera moves. It also discusses background elements. You can direct a scene. You don’t need a whole storyboard in one sentence.
Use this as a starting template for a speaker standing in a studio:
Medium shot of the speaker facing camera. As the first sentence begins, the camera slowly moves closer to a chest-up framing. The speaker keeps their shoulders level and speaks to the lens. Hold the final framing while the last line finishes.
It has an endpoint. Good. If the image is already an extreme close-up, though, asking for a chest-up view requires the model to invent a torso it can’t see. A reference image with room around the subject makes this direction more plausible. That’s a shot-design judgment, not an official image requirement.
Here’s a quieter alternative for a line that needs space rather than urgency:
Start in a steady chest-up shot. Keep the camera still through the first half of the audio. As the speaker reaches the final sentence, ease back slightly to show more of the studio. No sudden pan or cut.
Don’t stack a push-in plus an orbit, never mind a fast zoom just because the text box lets you. For a talking-head explanation, one move is usually plenty. The audio already has a rhythm; the camera should have a reason to interrupt it.
Write gestures where a viewer could see them
The ByteDance prompt guide linked from the product page gives an example of a person walking forward, standing with hands on hips before speaking and turning around. It also says the model can react to meaning in the audio without a text prompt. That means a gesture prompt should clarify your intended blocking instead of trying to transcribe every syllable into a hand movement.
For a presenter introducing a product, try:
The speaker faces the camera in a medium shot. At the phrase introducing the product, they open one hand toward the empty space on their right, then bring the hand back to rest. Keep the camera locked off and the face visible during speech.
Leave room. If you want to place a product shot beside the character in an edit, you need space for it. This prompt requests a gesture and framing; it doesn’t guarantee a clean composite or a usable hand. Look at the fingers and at the exact frame where the arm returns. If that transition fails, simplify the gesture before changing the voice track.
For a conversational line rather than a sales pitch:
The speaker listens for a beat, then looks back at the lens as the reply begins. They give one small nod after the first sentence and keep both hands relaxed. No large arm movements.
Don’t assign a nod to a nonexistent pause. If your audio starts mid-sentence, edit the pause into the audio or remove the listening beat from the prompt. A prompt can’t make a gap in an already continuous recording. Hear the gap first. Then write for it.
Treat background changes as their own experiment
BytePlus’s product page shows examples of adding elements not present in the reference image through prompts. The linked prompt guide also mentions control over background elements. Neither statement means a newly invented prop will land exactly where your editor needs it.
A restrained scene prompt is easier to inspect:
Keep the person centered in a warm, uncluttered recording room. A single practical lamp glows softly in the back corner. The speaker remains in the same position and the camera stays still for the whole clip.
Here you’re testing the room, not the camera. If the lamp drifts, you know which part of the request to remove. By contrast, “walk to a window, pick up a cup, spin toward a second person, then change the lighting” gives you several possible failure points and no clean diagnosis. Split that scene into separate clips if the edit permits it.
A two-person scene needs an even firmer assignment. The product page lists multi-person scenes, and the prompt guide says you can specify which person speaks. A draft instruction could read:
The person on the left speaks while the person on the right listens and looks toward them. Keep both faces in a steady two-shot. The listener stays silent and doesn’t turn toward the camera.
If the source image doesn’t clearly put two people in those positions, “left” and “right” are fragile labels. Fix the reference. First. Don’t claim the prompt can identify named individuals or assign separate voices unless your own workflow and current documentation support that claim.
Budget for one controlled revision
The BytePlus pricing panel lists $0.07 per second for its film-grade tier with camera and body movement, and shows 30 to 300 seconds of video output for that tier. A 30-second film-grade clip is 30 s × $0.07 = $2.10 at the listed rate. One first pass plus one full-length retry would be 2 × $2.10 = $4.20, before any other charges or account-specific terms. Check live pricing before you pay.
The same pricing panel separately lists $0.12 per second for an avatar tier focused on lip sync. Don’t apply the film-grade length statement to that tier without checking its own terms. For an overview of the model and access route, start at the OmniHuman 2 independent guide.
My rule for paid iteration: change one instruction per retry. If the first output gets the face and timing right but the camera moves too much, leave the audio and reference alone; replace the camera line with “fixed medium shot.” If you change all three inputs at once, you might spend another $2.10 and still not know what fixed it.
Review the seconds that actually matter
Watch the first second, the instructed move and the final held frame. Pause at the mouth during a plosive, at the visible hand during a gesture, and when the camera reaches its new position. Those are review targets. No output is guaranteed to fail there.
Keep a small record beside each attempt: source image filename and audio filename; full prompt and selected tier; clip duration and what you changed. If you can’t tell which render came from which instruction, you’ll accidentally compare different inputs and learn nothing. Name the shot after the action, such as “studio-fixed-gesture,” rather than after a vague adjective like “cinematic.”
When the last line needs a product overlay, leave clean space on the side you plan to use and request a held final frame. Then verify that exact frame before building the edit around it. A pretty opening shot won’t rescue an unusable ending.
