AI dance video prompts: how to write one that works

A good AI dance video prompt is one sentence with five parts: who is dancing, what single action they perform, where the camera sits, how the light behaves, and an instruction to keep the face, hair and body as supplied. Everything past that is decoration the model has to fit into a twelve second clip, and the more decoration you add, the less motion budget is left for the dance. This guide explains how the four studio looks map onto real prompt text, gives you eleven prompts you can copy and run in the studio, and lists the words that reliably break a render.

The four looks are prompts you can read and edit

The four looks in the studio, Classic bounce, Hype, Smooth groove and Cute, are not filters and they are not hidden model tricks. Each one is a short written direction, in the same text field you type into, that sets the energy, the camera behaviour and the lighting baseline before your own words are applied. That means you can read what a look is doing by watching what it changes, and you can override any part of it by writing the opposite.

In practice this gives you two ways to work. You can select a look and add a single correcting clause, which is the fastest route to a usable clip. Or you can select the closest look and then rewrite the camera and the action yourself, which is how you get something the four looks do not offer on their own, such as a final freeze, a specific arm gesture, or a slow pivot instead of a bounce. If you empty the prompt text and generate, the look carries the whole direction, which is a useful baseline to compare future attempts against.

One warning about combining: if your text contradicts the look, the text wins, and the result can look muddy. Do not select Smooth groove and then ask for fast sharp hits at the same time. Pick the energy first, then refine.

The anatomy of a usable dance prompt

Five parts, in this order, is enough for almost every clip.

• Subject. One person, or one animal, described briefly if you are in text mode and named as the person from the photo if you are in photo mode. Two subjects split the render.
• Action. One verb phrase, one repetition, one beat. "Steps side to side on the beat" works. "Steps, spins, waves and jumps" does not.
• Camera. Either locked off, or a single slow move such as a push-in or a drift to one side. A camera that changes direction mid clip has to invent geometry, and faces distort when it does.
• Light. One direction and one quality, for example soft window light from the left, or a single hard key from camera right. The render inherits whatever mood you name.
• Identity lock. The clause keep the face, hair and body as supplied, which tells the model to hold what the photo gave it rather than redesigning the subject.

Duration is the sixth element even though it is set by a slider rather than typed: a one beat action wants eight seconds, a full groove wants twelve. The studio offers 5, 8, 10 and 12 second clips, and the default render is Standard, 720p, 12 seconds, for 10 credits.

What to leave out

Leave out the song, the choreography name, the outfit changes, the crowd, the transitions and anything but one location. The output is a new AI performance rather than a reproduction of any recorded routine, so naming a specific video never helps. Leave out multiple camera angles for the same reason.

A copy-ready prompt template

Fill the blanks, keep the order, and keep it to one sentence where you can.

the [subject] performs [one action] on the beat, camera [locked off / pushes in slowly / drifts to one side], [light description], [length] seconds, [frame], keep the face, hair and body as supplied

And a worked version of the same template, ready to paste into photo mode:

the person from the photo performs a slow two-step bounce on the beat, camera locked off at chest height, soft window light from the left, 12 seconds, 9:16 vertical, keep the face, hair and body as supplied

Eleven prompts you can copy

The first two in each group are photo mode, meaning the subject is the person in your uploaded image. The third in each group is text mode, where the subject is described from scratch. Swap the action, keep the structure.

Classic bounce

the person from the photo performs a steady two-step bounce on the beat, camera locked off at chest height, soft even light, 12 seconds, 9:16, keep the face, hair and body as supplied

the person from the photo shifts weight side to side in a grounded bounce, camera locked off slightly wide, warm lamp light from the right, 10 seconds, 9:16, keep the face, hair and body as supplied

one dancer in a plain grey tee and jeans does a relaxed two-step bounce on an empty street at dusk, camera locked off wide, soft overhead light, 12 seconds, 16:9

Hype

the person from the photo throws one arm up once on the beat, camera pushes in slowly, hard rim light from behind, 8 seconds, 9:16, keep the face, hair and body as supplied

the person from the photo hits two sharp shoulder movements on the beat, camera locked off close, single hard light from camera left, 8 seconds, 9:16, keep the face, hair and body as supplied

one dancer in a black hoodie lands two sharp hits with the shoulders in an empty concrete room, camera pushes in slowly, one hard light from camera left, 8 seconds, 16:9

Smooth groove

the person from the photo rolls through a slow groove with the weight shifting side to side, camera drifts to the left at a walking pace, warm soft light, 12 seconds, 9:16, keep the face, hair and body as supplied

the person from the photo turns slowly in place while stepping through a smooth groove, camera locked off, dim warm practical light, 12 seconds, 1:1, keep the face, hair and body as supplied

one dancer in a silk shirt moves through a slow groove in a dim room lit by warm lamps, camera drifts to the left at a walking pace, 12 seconds, 16:9

Cute

the person from the photo does a small bouncy step with both hands at chest height, camera static and slightly high, bright flat light, 10 seconds, 9:16, keep the face, hair and body as supplied

one dancer in a pastel sweatshirt does a light bouncy step in a bright white room, camera static at eye level, flat bright light, 10 seconds, 1:1

One more for pets, which run through the same field with a different identity clause: the dog from the photo taps both front paws in a rhythmic bounce on the spot, camera low and static, soft daylight, 10 seconds, keep the fur, face and markings as supplied. The differences between animal and human renders are covered in the pets guide.

Intent to prompt fragment

When the clip comes back almost right, the fix is usually one fragment rather than a whole new prompt. Match the row to the problem and add the phrase on the right.

| What you want | Fragment to add | | --------------------------------- | --------------------------------------------------------- | | Steadier, more grounded energy | slow two-step bounce on the beat, camera locked off | | One clear big hit | throws one arm up once on the beat | | Slower, flowing movement | rolls through a slow groove, weight shifting side to side | | Short and playful steps | small bouncy step with both hands at chest height | | A final pose | holds the last pose for two seconds at the end | | Stronger resemblance to the photo | keep the face, hair and body as supplied | | Cleaner face in motion | bright even light, no shadow across the face | | A feed-ready crop | 9:16 vertical frame, subject centred |

Failure words and why they break a render

Some words look useful and are not. These are the ones worth deleting on sight.

Crowds and audiences. "Crowd cheering behind them" tells the model to fill the frame with bodies, which costs motion budget and adds faces that were never in your photo. Empty the frame instead, or leave the background to the existing environment.

Cuts and transitions. "Cut to a close-up", "then the camera swings around", "multiple angles": a twelve second clip has no room for editing, and asking for a cut usually produces a jump in the middle of the movement or a warped frame at the seam. One shot, one camera.

Two people. "Two friends dancing" makes the render choose, allocate motion to both and often produce merged or duplicated limbs. Render one subject per clip and cut them together afterwards, which also keeps the credits predictable because each clip is charged on its own spec.

On-screen text. "Lyrics appear on screen" or "add the title at the end" pushes the model to paint lettering into the pixels, which comes out as smeared shapes. Add captions in your editing app after the download, not in the prompt.

Named routines and named people. Do not name another person's choreography, a specific video, or a real public figure. The model produces a new performance, it does not reproduce a recorded one, and it will not swap a real person's face onto your body. Describe the movement in plain words instead.

Contradictory camera instructions. "Locked off but also zooming in and out" leaves the model to average two incompatible moves. Choose one.

How to fix a render that drifts

Drift means the output slowly stops resembling the input: the face softens, the clothing changes colour, or the subject ends the clip in a pose nothing asked for.

The cause is almost always that the clip is doing too much with too little held constant. Fix it in this order.

1. Shorten the clip. An eight second render has fewer frames in which to drift than a twelve second one.
2. Reduce the action to one verb phrase, and give it the full duration instead of dividing the clip between three movements.
3. Add or strengthen the identity lock clause, keeping the face, hair and body as supplied.
4. Firm up the light instruction, because a vague light lets the model reinterpret the whole scene halfway through.
5. If it still drifts, the source photo is the limit. A sharper image, or one where the face fills more of the frame, gives the model a stronger constraint to hold, as described in the photo guide.

Change one of those per attempt rather than all five, and test at 480p for 5 credits before spending a 1080p rate of 60 credits on the same idea. Which render length and frame to choose for a given feed is set out in the step-by-step walkthrough.

Reusing a prompt on a new photo

A prompt that worked on one photo is a reasonable starting point on another, but expect to adjust the light clause. If the new photo is lit from the opposite side, carry that into the prompt or drop the light instruction and let the look decide. The two model inputs, text and image, are independent, so the words never need to describe what the photo already shows. The technical background on how the model holds an identity across a clip is in the Seedance 2 explainer.

Frequently asked questions

How long should a dance prompt be?

One sentence is the target, and five clauses are plenty: subject, one action, camera, light and the identity lock. Text mode accepts 10 to 2,000 characters, so the field is forgiving, but a longer prompt does not produce a better clip; it usually spreads the motion budget thinner.

Do the four looks count as prompts?

Yes. Classic bounce, Hype, Smooth groove and Cute each set the energy, camera feel and lighting baseline in the same text field you type into, so you can select one and add a single correcting clause rather than writing a full direction from nothing.

Should I write negative instructions?

Avoid them where you can. "No crowds" still puts the idea of a crowd into the prompt, and the model may paint in a partial background crowd anyway. Describe the empty scene you want instead: an empty street, a plain wall, a bare room.

Why does my render ignore the look I picked?

Usually because the prompt text contradicts it. If you select Smooth groove and then ask for fast, sharp hits with a moving camera, the text overrides the look. Empty the text field, generate once as a baseline, and add your changes back one at a time.

Can I get a specific person's dance from a video?

No. There is no video input and no motion transfer, and no face swapping onto a real person. Every render is a genuinely new AI performance built from your photo or your description, so a prompt is a direction rather than a copy of anything recorded.

What is the cheapest way to test a prompt?

Render a 5 second clip at 480p for 5 credits. Once the photo and the words read well at that price, re-render the winner at 720p for 10 credits, and only then consider 1080p at 60 credits. The pricing page shows what each plan and one-time pack grants.