How AI Dance Face-Swap Videos Are Actually Made
The clips where someone performs choreography they never learned are motion transfer, not face swap. The pipeline in the order it runs, what it costs per second in August 2026, and why most apps selling it are the same endpoint underneath.
They are made with motion transfer, and in most cases nothing is swapped. A reference dance clip supplies the motion path, one or more photos supply the person, and a video model re-performs the choreography as that person. The face is not pasted onto a body — the whole body is regenerated frame by frame along the movement. That distinction explains almost every result people find confusing, including why wardrobe sometimes changes and why a cropped photo ruins the take.
The pipeline, in the order it runs
1. Pick the source clip. Five to ten seconds, one dancer, upright and repeating movement, full body in frame. Rapid direction changes and floor work transfer badly. The section that transfers best is usually the section viewers recognise anyway.
2. Prepare the character photo. Full body, feet visible, standing, neutral pose, plain background. None of that is what an image model or a phone camera gives you by default, and every one of those properties is load-bearing in the next step.
3. Run the motion transfer. fal-ai/kling-video/v3/pro/motion-control at $0.168 per output second, or the Standard tier at $0.126 — both checked 10 August 2026. It accepts a 3–30 second reference video plus a character image. Set character_orientation to video: on that setting the reference clip drives framing and the face-binding element works, and on image the binding is a silent no-op you still pay for.
4. Put the audio back. The trending track is half the format. Re-attach the original soundtrack rather than letting the model invent one.
An eight-second clip on the Pro tier is about $1.34. That is the whole cost of the format.
Why do all the dance apps look the same?
Because very few models ingest motion at all. Most consumer "AI dance" apps are a front end over the same one or two endpoints, priced in credits so the per-second cost is not visible. That is not an accusation — the interface, the queue, the upload handling and the rights check are real work, and plenty of people would rather pay for them than call an API.
It does give you a test, though. Divide what an app charges by the seconds of video it returns. If the answer is far outside the $0.126–$0.168 band, you are paying mostly for the wrapper. Our comparison of the models that can do this lists which ones ingest motion at all.
Why yours looks wrong
Four causes, roughly in order of frequency:
- The photo was cropped. A dance uses legs. Frame your character at the chest and the model invents the lower half — and the invention is exactly where the eye goes.
- The choreography is too busy. Simpler movement reconciles better against a body whose proportions do not match the source dancer's.
- The clip is too long. Drift compounds with duration multiplied by motion amplitude, and dance maximises the second term. Five to eight seconds.
- Wrong orientation setting. See step 3. It is the most common single cause.
The longer diagnostic goes through these with fixes.
Doing it without assembling the pipeline
If you do not want to manage endpoints and parameters, our AI dance video generator is this exact pipeline behind two inputs — a reference video and one photo of you — returning a vertical MP4. Entry is $0.99 in credits, no subscription.
The rights part, which is not optional here
Three separate claims sit in one dance clip: the choreography, the soundtrack and the likeness of the person performing it. Any tool that copies movement out of someone else's video should ask you to confirm you have the right to use that source before it runs — ours does. Use your own face, and check the audio licence before you post.