Why Your AI Dance Video Looks Wrong (Five Causes, In Order)

You picked a dance, supplied a photo, paid, and what came back is subtly not you — or is you, doing something a body cannot do. Work through these in order. They are sorted by how often each turns out to be the cause.

1. The orientation parameter

On Kling 3.0 Pro Motion Control, character_orientation has to be video.

In that mode the reference clip drives the framing and the model honours the element that binds your face to the character. On image, the framing follows your still and that face binding does nothing at all. The request succeeds, you are billed, and the identity drifts.

This single setting accounts for more "motion transfer does not work" reports than every other cause combined. It also gates the 30-second reference ceiling.

2. The source clip was never going to transfer

Not every dance survives. What breaks it:

• Two people in frame. The model has to decide which body is yours.
• Rapid cuts and whip pans. Contradictory motion to follow.
• Cropped limbs. Anything outside the frame gets invented, and the invention is where the eye goes.
• Floor work and fast direction changes. Reconciling those against different limb proportions is the hardest case there is.

Pick five to eight seconds of clear, upright, repeating movement. That is also the section viewers recognise, so you lose nothing.

3. Not enough reference angles

A dance turns your head through angles a selfie never covers. One frontal photo tells the model nothing about your profile, so it invents one, and the invention does not match.

Supply frontal, both three-quarters and a profile, in similar lighting, whole head in frame. If wardrobe drifts too, shoot the references in the outfit — reference-driven models read clothing from your photos, not from the source dancer.

4. The clip is too long

Drift is duration multiplied by motion amplitude, and dance maximises the second term. Eight seconds holds; fifteen slides. For a longer piece, generate two shots and cut between them.

5. Only now, the tier

Standard motion control is cheaper than Pro, and in our side-by-side runs Pro held identity visibly better. At $0.168 per second, an eight-second clip on Pro is $1.34 — a third of a dollar more than Standard to avoid a re-run.

Prices checked against the live endpoints on 3 August 2026.

What does not help

Re-rolling the same request. If the cause is structural — wrong orientation, one reference photo, a bad source clip — the next generation fails identically and costs the same again. Change an input first.

Upscaling. Drift is temporal, not spatial. A sharper render shows it more clearly. If the clip is otherwise good and only soft, that is a different problem — see Topaz Video Upscale at $0.01 per second, which is a finishing step and not a fix for identity.

Switching to a reference model. HappyHorse 1.1 Reference to Video and MiniMax H3 Reference to Video build a person from photos but never ingest a source video's motion. They will give you a dance, not that dance.

The two-minute checklist

Orientation video. Five to eight seconds. Upright, single-person source. Four reference angles in the right outfit. Pro tier if the face is on screen. That combination works far more often than it fails.