OmniHuman 2 Prompt Guide: Direct Gestures, Emotion and Camera

How to structure an OmniHuman 2 prompt

Think in three slots: performance, gesture and camera. Performance is the overall mood (“warm and confident”, “playful”, “serious newsreader”). Gesture is one or two concrete actions (“counts on fingers”, “points at the viewer”, “holds the jar up to the lens”). Camera is a single move or framing (“static medium shot”, “slow push-in”, “gentle orbit to the left”).

Keep it under about 40 words. Long prompts with many competing actions tend to produce busy, unnatural motion; it is better to render twice with two focused prompts than once with an overloaded one.

Camera directions

Use one move per clip. These phrases are clear and widely understood:

• “static medium shot” — calm, presenter style; best for long explainers.
• “slow push-in” — builds intensity; great for ad hooks and emotional lines.
• “gentle orbit around the subject” — adds production value to songs.
• “handheld feel, slight sway” — makes UGC-style clips look phone-shot.
• “wide shot, then slow zoom on the speaker” — useful in group scenes.

Gestures and body language

Name the gesture and, if it matters, when it happens: “raises an eyebrow at the question”, “opens both hands on ‘everyone’”, “nods while listening”. Product shots work best when the product is already in the photo; the prompt can then ask the person to lift, tilt or point at it.

For singing, describe breath and movement rather than mouth shapes — “closes eyes on long notes, sways with the beat, hand on chest in the chorus” — and let the audio drive the lips.

Prompt examples by format

Swap the details for your own photo and audio.

1. 1UGC skincare hookexcited, talks fast, holds the bottle close to the camera on the first line, handheld feel
2. 2Founder announcementcalm and confident, open hand gestures, static medium shot, small smile at the end
3. 3Balladeyes closed on long notes, gentle sway, slow orbit, soft emotional expression
4. 4Two-host podcastleft host speaks, right host nods and reacts, static wide shot, relaxed conversation
5. 5Lessonteacher tone, counts points on fingers, glances down as if at notes, static shot
6. 6Mascot greetingcheerful, waves at the start, bounces slightly, slow push-in

Languages and troubleshooting

You can write direction in English, Chinese, Japanese, Korean, Spanish (Mexico) or Indonesian. The spoken audio can be in any language — lip sync follows the sound, not the prompt.

If motion looks exaggerated, remove adjectives like “very” and “dramatic”. If nothing seems to happen, check that the gesture is physically possible in the framing: hands can’t gesture if the photo is a tight face crop.

Common prompting mistakes

The most frequent mistake is describing the script instead of the performance. Writing “she explains our new pricing” adds nothing, because the audio already carries the words; “leans forward and taps the table on each price” gives the model something to do.

The second is contradiction: asking for a static shot and an orbit, or for calm delivery and energetic jumping, in the same clip. When instructions fight, the motion becomes muddy. Decide on one mood and one camera behaviour per render.

• Avoid emotional extremes unless the audio matches them — a whispered line with “shouting angrily” looks broken.
• Don’t ask for objects that aren’t in the photo to appear; add them to the image instead.
• Skip technical film jargon the model has no reason to know, such as specific lens brands.

Iterating without wasting credits

Lock the photo and audio first, then vary only the prompt between test renders. Run a five-second excerpt of the most important line with two or three prompt variants side by side — at 50 credits each, that comparison costs less than one long render that misses.

Once a prompt works, save it with the project as a reusable “direction preset” for that presenter. Series look more consistent when every episode starts from the same direction text.