How-to
A prompt formula for viral AI dance videos
A five-part prompt template for AI dance videos from a single photo, the rules that decide whether a model follows it, the prompt behind our salsa clip, four sample dances, and how to use your own reference clip.
Published November 6, 2026

A prompt for a viral-style AI dance video needs five parts: the dance, the framing, a timed list of moves, an ending pose and a line about the music. Below are the template, the rules that decide whether a model follows it, and the prompt behind one of our own clips. If you would rather skip the writing, The Dance effect makes the clip from one photo and a dance you pick.
Upload one full-body photo, pick a dance, and generate. No prompt needed.
The five-part template
We wrote and tested this for image-to-video with a single person in frame. The photo carries the look, so the prompt only describes the choreography.
[Dance style], full body in frame, camera locked static, vertical. [0-2s] Main move 1, plus one or two small supporting moves. [2-5s] Main move 2. [5-8s] Main move 3. [8-10s] Ending freeze: the pose, the expression, where the eyes go. Fast and precise, every move snapping on the beat, smooth continuous natural motion. AUDIO: one line describing the music.
- The first line sets the dance, the framing and the camera. Keep to one dance per clip, since mixed styles made the moves break and stutter.
- The timed lines carry the choreography. Each segment holds one main action, so there is no blur of moves for the model to average.
- The ending freeze gives the clip a last beat and a place for the expression and the eyes.
- The quality line pushes speed and precision, because models soften the energy they are given.
- The audio line only matters if your model makes sound in the same pass. The Dance clips do, and for a dance clip the music is part of the result.
The rules that decide whether it works
- Segment the timeline. Our first long, flowing prompt came back with almost no dance in it. Rewritten as timed segments with one main action each, the same model started dancing.
- Keep the prompt short. The photo already shows the person, the outfit and the room, so do not describe them again. Long identity text takes space the moves need.
- Write physical verbs and push them. Snapping, bouncing, popping, flicking and shimmying are reliable. Models discount intensity, so "hips swaying" works better as "hips snapping hard on every beat".
- If a segment is ignored, it probably holds two main actions. Split it, or cut one. Adding adjectives does not help.
- Write turns as a half-turn and a glance back. A full spin can break the face. The red-dress clip below keeps one full spin, and the flaring skirt is what makes it the highlight.
Choose the input photo like you choose the prompt
- Show the whole body, feet included, with the person clear and large in frame.
- Avoid mirrors in the scene. A reflection in a video falls apart.
- Fitted clothes are steadier than oversized ones. Loose fabric tends to interfere with how the model tracks the joints.
- A flowing skirt or long hem is a bonus. Give it one beat in the prompt, such as a flick or a turn, and it becomes the highlight of the clip.
Two ways to write it
Hand-written sequences give you control over the density of moves, which suits something like a fast salsa. Naming a dance the model already knows works differently. In our tests a well-known routine such as the Macarena or Renegade came out more fluid and more convincing than anything we wrote move by move, because the model draws on a real choreography. Pick a named dance for a famous routine and a written sequence when you want to design the beats.
The person dances the [name of the dance] in the current scene, full body in frame, camera locked static. No singing.
This is the prompt behind our salsa clip, a hand-written sequence for the red-dress photo. It follows the template above, and the skirt flick and spin sit in the second segment.
She dances an up-tempo salsa TikTok routine, full body in frame, camera locked static, vertical. [0-2s] Hips snapping hard side to side on every beat, quick salsa step-taps, shoulders shimmying fast. [2-5s] She grabs the long red skirt and flicks it out wide, then one fast full spin, the dress flaring around her legs, landing facing the camera right on the beat. [5-8s] Rapid hip shakes with both arms raised high, then a sharp hair toss. [8-10s] She snaps into a freeze: one hip popped hard, chin down, eyes locked on the lens, confident grin. Fast, precise, every move popping exactly on the beat, smooth continuous natural motion throughout. AUDIO: up-tempo Latin pop with heavy percussion and a driving beat.
Dancing to your own choreography
If you already have the dance, use Your Move on the Dance page. You upload your photo and a short clip of the routine, and the moves from the clip transfer onto the person in the photo. Reference clips run from 3 to 15 seconds, the result runs for the length of the clip, and the clip's own audio stays. The price follows the clip's length in seconds, and the panel shows it before you generate.
- Use a clip with one dancer, head, shoulders and torso clearly in frame throughout. If the engine cannot find a clear person it refuses the clip and your credits come back.
- Use a clip you have the rights to. A clip the engine flags as copyright-protected is refused, so filming a short routine yourself avoids that.
- Pick a photo of one person, ideally full body. The photo's shape must sit between 2:5 and 5:2. Anything taller or wider is stopped before a single credit is spent.
- On the Dance page, upload the photo, switch to the tab for your own video and add the clip.
Crowd-dance trends such as Storm II ask for a lead dancer with a crowd behind. reRAW makes the lead dancer this way. It does not make the crowd, and we have not tested a crowd version.
Questions
Do I have to write a prompt?
Which photo works best?
Does the video have music?
Can I use a dance of my own?
Why did my dancer's face change during a spin?
Try a dance with your own photo.