On this page
- Short answer
- The OmniHuman 2.0 prompt template
- 12 OmniHuman 2.0 prompts to copy
- The template, part by part
- What a take costs
- How OmniHuman 2.0 reads a prompt
- Four rules for every prompt
- Do and don't
- Prompts with several characters
- Sequences: first ... then ...
- Singing and rap
- Get the inputs right first
- Consent and responsible use
- Frequently asked questions
Short answer
In OmniHuman 2.0 the audio already decides the words, so the prompt only directs the performance. Write one or two natural sentences in this order: camera movement, the speaker's emotion, whether they speak, sing or cry, their actions, then any background events or other characters. Describe motion only, never what the photo already shows, and say who speaks. The prompt is optional and can be up to 2,000 characters.
Get started
The OmniHuman 2.0 prompt template
- 1
Camera movement
Start with the shot and how it moves: static medium shot, slow push-in, gentle handheld feel, slow orbit, low angle. One movement per take is plenty.
- 2
The speaker's emotion
Say how they feel while they talk: warmly, calmly, excitedly, grumpily, seriously. Match the tone of the voice track, or the face and the voice will disagree.
- 3
Speaking state
Say what the mouth is doing: "the man speaks", "she sings", "he raps", "she speaks while crying". Write it explicitly, with the subject, every time.
- 4
Specific actions
Add two or three concrete actions: nods on key points, lifts the bottle towards the camera, counts on her fingers, waves at the end.
- 5
Background and other characters
Optional: what happens around the speaker (petals drift past, neon signs flicker) and what anyone else in the frame does (the man beside her listens and nods).
12 OmniHuman 2.0 prompts to copy
Grouped by use case. Each card says why it works and suggests a length and resolution, with credits from the studio's Lip Sync price (4 per second at 720p, 6 at 1080p). Change the subject words (the woman, the chef, the cat) to match your photo. Text only: we never pair a prompt with a clip it did not make.
Ads, presenters and product videos
A clear speaker, a few purposeful gestures, often a product in hand.
Product presenter
Slow push-in from a medium shot. The woman speaks with a bright, confident smile, lifts the bottle in her right hand towards the camera when she says the product name, then sets it down and gestures to the side as she lists the benefits.
Why it works: The gesture is tied to a moment in the script (the product name), so it lands where it matters.
- OmniHuman 2.0
- 20 seconds of audio
- 1080p
- 120 credits
Calm spokesperson
Static camera at eye level. The man speaks calmly and sincerely, leaning slightly forward on the important lines. First his hands rest together on the desk, then he opens them towards the camera as he finishes.
Why it works: A locked camera and small, slow movements read as trustworthy for a founder message or an apology.
- OmniHuman 2.0
- 30 seconds of audio
- 1080p
- 180 credits
Courses, training and updates
Calm delivery and steady framing, so viewers listen rather than watch the camera.
Course explainer
Static medium shot. The teacher speaks patiently and clearly, counting the points on her fingers, then turns slightly towards the whiteboard behind her and gestures at it before looking back at the camera.
Why it works: Counting on fingers and one turn to the board give a long lesson some movement without distracting from the voice.
- OmniHuman 2.0
- 1 minute of audio
- 720p
- 240 credits
Friendly team update
Gentle handheld feel, medium close-up. The manager speaks in a relaxed, friendly way, nodding now and then, and gives a small thumbs up at the end.
Why it works: A slight handheld feel makes an internal message look less staged than a locked-off shot.
- OmniHuman 2.0
- 45 seconds of audio
- 720p
- 180 credits
Singing and rap
Say "sings" or "raps", and let the camera and body move with the track.
Slow ballad
Slow orbit around the singer from left to right. She sings softly with her eyes closed, then opens them and looks into the camera when the chorus starts, one hand pressed to her chest.
Why it works: "Sings" tells the model this is a song, and the change at the chorus follows the structure of the track.
- OmniHuman 2.0
- 1 minute of audio
- 1080p
- 360 credits
Rap verse
Low-angle handheld shot that moves with the beat. The rapper performs the verse with energy, chopping his right hand in rhythm and leaning into the lens on the last line. Neon signs flicker behind him.
Why it works: Fast rap needs a clean vocal; the prompt adds energy and one background event without describing the outfit already in the photo.
- OmniHuman 2.0
- 30 seconds of audio
- 1080p
- 180 credits
Animals, cartoons and characters
The same template works for a cat, an anime girl or a 3D mascot.
Talking cat
Static close-up. The ginger cat speaks in a grumpy, dramatic way, narrowing its eyes and flicking its ears, then turns its head away as if offended.
Why it works: Animal-specific actions (ears, eyes, head turn) make a pet feel alive instead of just moving its mouth.
- OmniHuman 2.0
- 10 seconds of audio
- 720p
- 40 credits
Anime character
Fixed camera, medium shot. The anime girl speaks excitedly with wide eyes, waves with both hands and laughs at the end. Cherry blossom petals drift past in the wind.
Why it works: The drifting petals are a background event, the kind of thing the photo cannot show on its own.
- OmniHuman 2.0
- 10 seconds of audio
- 720p
- 40 credits
Several characters and sequences
Name who speaks and who reacts, and use first ... then ... for changes over time.
One speaks, one listens
Static two-shot. The woman on the left speaks with a serious expression while the man on the right listens, nods slowly and raises his eyebrows in surprise at the end. Only the woman speaks.
Why it works: Positions (left, right) say who is who, and "only the woman speaks" stops the listener from mouthing the words.
- OmniHuman 2.0
- 15 seconds of audio
- 1080p
- 90 credits
A change of mood
Slow push-in. First the old man listens quietly with a faint smile, then he speaks warmly to the camera, and finally he laughs and leans back in his armchair.
Why it works: First, then and finally give the model an order of events that matches the pauses in the audio.
- OmniHuman 2.0
- 20 seconds of audio
- 720p
- 80 credits
Emotional monologue
Static close-up. The young man speaks quietly while crying, wiping his cheek with the back of his hand, then takes a breath and manages a small smile.
Why it works: Crying is a speaking state of its own; naming it keeps the face consistent with a shaky voice.
- OmniHuman 2.0
- 20 seconds of audio
- 1080p
- 120 credits
The template, part by part
Put the parts in this order and write them as one or two sentences, not as a list.
| Part | What to write | Example |
|---|---|---|
| Camera | Shot size and one movement | Slow push-in from a medium shot. |
| Emotion | How the speaker feels | The woman speaks with a bright, confident smile, |
| Speaking state | Speaks, sings, raps or cries, with the subject named | (already in the line above: the woman speaks) |
| Actions | Two or three concrete movements, in order | lifts the bottle towards the camera, then sets it down and gestures to the side. |
| Around them | Background events and other characters' reactions | Steam rises from the cup beside her. |
What a take costs
Studio credits for Lip Sync, by length of the voice track. The prompt does not change the price.
| Audio length | 720p | 1080p |
|---|---|---|
| 4 seconds | 16 credits | 24 credits |
| 10 seconds | 40 credits | 60 credits |
| 30 seconds | 120 credits | 180 credits |
| 1 minute | 240 credits | 360 credits |
| 5 minutes | 1,200 credits | 1,800 credits |
Billed per started second, minimum 4 seconds. Lip Sync is available on any paid plan or credit pack, and credits for a video that fails come back automatically.
How OmniHuman 2.0 reads a prompt
OmniHuman 2.0 is ByteDance's audio-driven video model, and the audio comes first. The words, their timing, the intonation and much of the emotion are already in the voice track, and the model turns them into mouth shapes, facial expression and body movement on its own. That is why a Lip Sync video works with no prompt at all.
The prompt is for what the audio cannot say: where the camera goes, what the hands do, how the mood changes, what happens behind the speaker, and how anyone else in the frame reacts. Think of yourself as a director talking to an actor who already knows the lines.
KreatorFlow is an official partner of BytePlus, a ByteDance company, and runs OmniHuman 2.0 through BytePlus, a ByteDance company. The advice on this page follows the guidance BytePlus publishes for the model, in our own words, with our own examples.
Four rules for every prompt
Write natural, coherent sentences. A keyword list ("smile, nod, cinematic, 4k") gives the model nothing to order, and a long story gives it too much to fit. One or two sentences describing a continuous performance is the sweet spot.
Describe motion and events only. The model can already see the photo, so "a woman with long blonde hair in a white T-shirt" adds nothing and can even fight the picture. Refer to the subject simply (the woman, the cat, the singer) and spend your words on what changes.
Be clear, consistent and specific. "She nods twice and smiles" is better than "she reacts naturally". Do not ask for opposite things at once, such as a static camera that also orbits, or a calm speaker who shouts.
Use as little negation as you can. Say what should happen rather than what should not. The one useful exception is in group shots: "only the woman speaks" is a positive way to keep the others quiet.
Do and don't
A quick checklist before you press Generate.
- Do name the speaker and the state: "the man speaks", "she sings".
- Do match the emotion in the prompt to the tone of the voice.
- Do give actions an order with first, then and finally.
- Do keep one camera movement per take.
- Don't describe clothes, hair or the setting already visible in the photo.
- Don't write a keyword list or a full story.
- Don't give conflicting instructions (static and orbiting, calm and shouting).
- Don't stack negatives ("no blinking, no head turns, no hands").
Prompts with several characters
OmniHuman 2.0 can animate a photo with more than one person or animal, where one speaks and the others react. The voice track belongs to one speaker, so the prompt has to say clearly who that is.
Point at the speaker by position or a feature that is unique in the photo (the woman on the left, the man in the red jacket, the grey cat), and give the listeners their own small actions: nodding, glancing at the speaker, raising an eyebrow. Close with "only the woman speaks" if a listener starts mouthing along.
Template for a two-person photo
Static two-shot. The [speaker, by position] speaks [emotion], [action]. The [listener, by position] listens, [reaction], and [reaction at the end]. Only the [speaker] speaks.
Sequences: first ... then ...
A Lip Sync take can last up to 5 minutes, and a long take needs more than one mood. Describe the changes in order with first, then and finally, and line them up with the structure of the audio: a pause, a change of topic, a chorus.
Keep it to three or four stages. More than that tends to blur, and in a long take the voice keeps the performance alive between the stages anyway. For a very long video, such as a full lesson, a calm base performance with one or two changes often looks more natural than constant action.
Sequence template
[Camera]. First the [subject] [state and action], then [state and action], and finally [state and action].
Singing and rap
Write "sings" or "raps", not "speaks", so the model treats the track as music. Then direct the performance as you would for a music video: a slow orbit or push-in for a ballad, a low handheld angle that moves with the beat for rap, eyes closed on soft lines and a look into the camera on the chorus.
The vocal matters more than the prompt. A clean vocal with little backing music gives the most accurate mouth movement, especially on fast rap. If you only have the full mix, choose a version where the voice sits clearly above the instruments. Use songs and vocals you have the rights to.
Get the inputs right first
No prompt can fix a weak photo or noisy audio. The best photos have a clear subject, an unobstructed face (no hands, microphones or hair across the mouth) and a stable composition, as a half-body or head-and-shoulders shot. The best audio is clean speech or vocals with little background noise.
The voice track can be 4 seconds to 5 minutes long, and the video is exactly as long as the audio. Test a short line at 720p before you render a long take at 1080p.
Consent and responsible use
Lip Sync makes the person in a photo say whatever the audio says, so use it only with permission. You must be the person in the photo or have their permission to make them speak, and you must have the right to use the voice and any music. On the website you confirm this the first time you generate with a new photo.
Never use it to deceive: no impersonation, no fake statements from real people, no harassment. Sexual content is never allowed, nor anything involving someone under 18, and photos with nudity or of minors are refused before anything is charged. When you share a realistic talking video of a real person, say that it is AI-generated. To report misuse, write to support@kreatorflow.ai.
Frequently asked questions
Do I need a prompt for OmniHuman 2.0?
No. The audio drives the mouth, expression and movement, so a Lip Sync video works without one. Add a prompt to direct the camera, gestures, mood changes, background events and other characters.
What is the best prompt structure for OmniHuman 2.0?
Camera movement, then the speaker's emotion, then whether they speak, sing or cry, then specific actions, then optional background events or other characters, written as one or two natural sentences.
Should I describe the person in the photo?
No. The model sees the photo, so describe only motion and events. Refer to the subject simply, such as the woman, the singer or the cat.
How do I make only one person speak in a group photo?
Name the speaker by position or a unique feature, give the others listening actions, and end with "only the woman speaks" (or whoever it is).
How do I prompt singing or rap?
Write "sings" or "raps" and direct it like a music video. Use a clean vocal track: it matters more than the prompt for accurate mouth movement.
How long can the prompt be?
Up to 2,000 characters in the KreatorFlow studio and API, but one or two sentences is usually best.
Does the prompt change the price?
No. Lip Sync is priced by the length of the audio: 4 credits per second at 720p and 6 at 1080p.
