You've seen the clips. A still photo of someone, or a character somebody drew, suddenly doing the dance that's everywhere that week. You try it yourself and get a figure whose legs melt into the floor by second four. Making an AI dance video from a photo has less to do with the prompt than you'd think. It comes down to two inputs agreeing with each other.
This guide uses Kling 2.6 Motion Control, which runs in the Video to Video tool in two tiers, Standard and Pro. Kling has since released a 3.0 version of Motion Control. We don't run that one, so everything below is about 2.6.
How an AI dance video from a photo works: one clip for motion, one image for the look
Motion control takes two files. A reference video supplies the movement: the steps, the arm swings, the head turns, the timing. A character image supplies what you see: the face, the outfit, the body shape. The model reads the pose in each frame of the video and redraws your character in that pose.
That split explains most of what you'll get. The model never invents choreography. If the reference dancer spins, your character spins. If the reference is shaky, blurry, or half off screen, the model is guessing, and guesses show up as wobbly hands and arms that swap sides.
There's an optional prompt on top. Kling's own tutorial says to use it for "production direction, not choreography", meaning the light, the style, the setting. Something like "neon-lit rooftop at night, light rain, cinematic" works. "She kicks left then spins twice" fights the reference instead of helping it.
Pick a reference clip the model can read
The reference clip decides more of the result than your photo does. Kling's tutorial asks for 3 to 30 seconds of motion with the full body clearly visible. A clip that passes these checks is far easier for the model:
- One dancer. A second person crossing the frame muddles which body to follow.
- No cuts.
- The whole body in frame the entire time, feet included.
- A steady camera and a background that barely moves.
- Moderate speed. Fast footwork and quick hand flicks blur, and blurred limbs become guessed limbs.
- Arms that don't stay wrapped around the body for long, since the model can't track what it can't see.
Trending dance clips often fail two or three of these at once: jump cuts on the beat, a phone camera swinging around, a friend walking into shot. You'll usually do better filming yourself or a friend doing the routine against a plain wall, in one take, with the phone propped on a shelf. It also settles the question of whose performance you're copying.
Match your photo to the clip
This is the rule people skip, and Kling puts it plainly:
If your video shows a full-body shot of a dancer, your image must also be a full-body shot.
A head-and-shoulders portrait paired with a full-body dance forces the model to invent legs, hips, and shoes it has never seen, and it invents them a little differently in every frame. That's where the melting comes from. Check the photo against the clip before you upload:
| Check | Why it matters |
|---|---|
| Same framing: full body with full body, half body with half body | Mismatched framing is the most common cause of warped bodies |
| Head, arms, and legs all visible | Hidden parts get invented, and invented parts drift |
| Similar camera angle | A photo shot from above next to an eye-level clip stretches proportions |
| Arms apart from the torso | Visible gaps help the model keep limbs separate |
| Sharp, well lit face | Soft detail in the photo turns into shimmer once it moves |
| Short edge at least 300px | Kling's stated minimum. Bigger is better |
If the photo you have doesn't fit, fix the photo instead of hoping. You can try extending a cropped portrait into a full-body shot, or changing an awkward pose, in Image to Image, then bring the result over. The cost is an extra generation and the risk that the edit shifts the face slightly, so compare the two side by side before you commit.
Orientation: follow the video or follow the image
The composer has a Character orientation setting with two choices. The name sounds technical. The difference isn't.
- Follows video is the default. Which way the character faces, its expressions, and the camera framing all come from the reference clip. If the dancer turns their back, so does your character. The studio accepts a reference clip up to 30 seconds in this mode.
- Follows image still takes the moves from the video, but keeps the character's orientation anchored to your photo, and lets you steer the camera with the prompt. The reference limit drops to 10 seconds.
For dance, pick Follows video. Dances turn, step sideways, and face away, and this mode lets the character do all of it. Follows image suits cases where the photo's pose matters more than matching every turn, like a front-facing mascot that should stay front-facing while it bobs to a beat. You pay for that with the shorter limit and some stiffness when the dancer rotates hard.
Standard vs Pro
Both tiers take the same inputs and the same settings. Pro is the pricier one, and it's aimed at higher fidelity: cleaner detail, and hands, faces, and fabric that hold together better through fast moves. Standard is cheaper and good enough to find out whether a photo and clip pair works at all.
Our recommendation: run the first attempt on Standard with a short cut of the reference, around 5 to 8 seconds. If the body tracks the dance and the face holds, run the full clip on Pro. If Standard falls apart, don't expect Pro to save it. A framing mismatch or a blurry reference breaks both tiers, so fix the inputs first.
In both tiers the cost scales with the length of the reference clip. The composer shows the credit estimate for your exact clip before you press Generate, so a 25 second run on Pro is something you choose on purpose, not something you find out about later.
Sound, length, and loops for social
Motion control keeps the reference clip's original sound by default. For a dance that's what you want, because the beat lines up with the moves: they came from the same video. There's a toggle in the composer to turn it off if you plan to add a track in an editor. If you do swap the music, keep the same tempo, or the steps will land off the beat.
Length is where most people overspend. A social dance post rarely needs 30 seconds. The recognizable part of a routine is usually a short phrase, so cut the reference down to that before you upload. Shorter clips cost less and leave less time for small errors to pile up.
For a loop, trim the reference so the dancer finishes in roughly the pose they started in. The output follows the reference pose by pose, so a clean loop in the source tends to give you a clean loop in the result. And if the post is for a phone screen, keep both inputs vertical so they agree on shape as well as framing.
Common failures and fixes
| What you see | Likely cause | Fix |
|---|---|---|
| Legs or feet melt or flicker | Photo cropped tighter than the reference | Use a full-body photo, or a reference framed like your photo |
| Arms swap sides or merge into the body | Limbs overlap the torso in the clip or the photo | Pick a clip with clear gaps between arms and body |
| Face changes partway through | Small or soft face in the photo, or long back-facing sections | Use a sharper photo with a larger face; skip clips where the dancer faces away for long |
| Character jitters on the spot | Handheld, shaky reference | Film on a tripod or a shelf |
| The wrong body gets tracked | Two people in the reference | One dancer only |
| Stiff, flat turns | Orientation set to Follows image | Switch to Follows video |
| Odd moves the dancer never made | Prompt describes choreography | Describe the setting and light, not the steps |
When a run fails, change one thing and rerun on Standard with a short clip. Swap the photo, the clip, and the prompt all at once and you learn nothing about which one was the problem.
The short version
- Motion control takes the moves from a video and the look from a photo. The two have to agree.
- Use a one-person, one-take reference with the whole body in frame and a steady camera.
- Full-body photo for a full-body dance.
- Follows video for dance, up to 30 seconds. Follows image when the photo's pose matters more, up to 10 seconds.
- Test on Standard with a short cut, then run the final on Pro.
- Only animate people who agreed to it.
Open Video to Video and try a short test with your own photo, or look through the other video tools first. If you want a still to move without a reference clip, Image to Video is the one for that.
