Short answer: don't describe the dance in a prompt. Give the video model a real dance clip to copy and one picture of your character in the opening pose. That method is called motion control. In our test on Kling 2.6 Pro, a 13.8-second trend dance came back with every beat in place and the character's face holding up to the final close-up. It cost about $1.55 and took just under 10 minutes. The part that decides whether it works is the clip you pick, not the model.
This is first-hand: we made a first dance this way on 4 November 2026, compared two models on 7 and 8 November, and then added it to the video options our AI team can choose from. Prices below were re-read on fal's model pages on 10 November 2026. Like our other guides, this one has no pictures of our influencers, only what we did and what we measured.
Copying a dance starts with picking the right clip.
Why prompting a dance doesn't work
A normal image-to-video model gets a start picture and a sentence. "She does the trending shoulder-shimmy dance" leaves it to guess the moves, the count and the timing, and it guesses wrong: you get swaying, a spin that wasn't asked for and hands that wander. A dance is a sequence of exact positions on exact beats, and words can't carry that.
Motion control models take a second input: a reference video. The model follows the body movement in that video and draws your character doing it, starting from your picture. On fal, Kling's motion-control endpoints take three things: your character image, the reference video and a short prompt.
The 5 steps we used
1. Pick a clip the model can read
This mattered more than anything else. fal's own page says the reference "should contain a realistic style character with entire body or upper body visible, including head, without obstruction". From our two tries, add:
One dancer, one continuous shot. No cuts, no second person walking in.
Facing the camera most of the time. Our first dance used a stock clip where the dancer looks away for much of it. The video moved fine, but the character's face never came through: the best single frame scored 0.30 on our face check, where we only keep frames at 0.45 or above. Our second test used a dancer facing the phone, and the face held.
Full body or at least the upper body in frame the whole time, plain room, steady phone.
Length: 3 to 30 seconds with the orientation set to follow the video (10 seconds if it follows your picture instead).
2. Draw the first frame in the clip's opening pose
Your character picture should match the first frame of the reference: same pose, same framing, same distance to the camera. We drew it with GPT Image 2.5 from three pictures: the character's reference sheet, her face reference, and the clip's first frame "only for her pose and the framing". We changed the room and the outfit to hers. If the opening pose doesn't match, the model has to jump from your pose into the dance in the first second, and that jump is where it looks fake.
3. Set the orientation to "video"
Kling's motion control has a character_orientation switch. fal's description: "video" matches the reference video and is "better for complex motions (max 30s)"; "image" matches your picture and is "better for following camera movements (max 10s)". For a dance, use "video".
4. Decide on the sound
The endpoint can keep the reference video's original sound. For our own pipeline we turn it off and add the sound on the platform instead (see the rules below): a trending sound added inside the app is the platform's to license, while a song baked into your file is yours to clear.
5. Check it at full size, frame by frame
Don't judge the result from a thumbnail. Step through it, and look hardest at the fast moves, the smile and the last second. Our own clip review takes about 2 stills per second of every clip and face-checks each one.
The reference clip carries the steps; the prompt can't.
What it cost and how long it took
Our two runs, same character picture, same 13.8-second clip:
Kling 2.6 Pro motion control
Kling 3 Pro motion control
Price on fal (10 Nov 2026)
$0.112 per second
$0.168 per second
Our 13.8 s dance
about $1.55
about $2.32
Time to render
9 min 45 s
about 7 min
Moves
every beat matched
the same, beat for beat
Hands in fast moves
a little smeared
a little sharper
Face
held, including the close-up
a tie with 2.6
The cheaper "standard" tiers are $0.07 per second (2.6) and $0.126 per second (3). We did not test them. Our earlier stock-clip dance, 10 seconds on 2.6 Pro, cost $1.10.
Our pick: Kling 2.6 Pro. Kling 3 Pro was 50% more per second for sharper hands and the same face, so 2.6 Pro is the motion option our AI team can choose today. To price a whole reel, the free AI video cost calculator does the sum.
What broke
The smile at about 11 seconds. Both models smeared it for a moment; Kling 3 drew something like braces on the teeth.
A tuft of hair on top of her head in the final close-up, in both versions. It came from the first frame we drew, not from the video model. Fix the picture before you animate; the video copies its mistakes.
Fast hands smear, the way they do in a real phone video. Fine at normal speed, visible when you pause.
A dancer who looks away costs you the face. The stock-clip dance moved well and failed the face check. That one video was the reason step 1 exists.
Check the fast moves and the last second at full size.
Which clips you can copy
Short version: copy a clip you have the right to use, and let the platform supply the music. This is not legal advice.
Your own recording is the safest reference: you, a friend or a dancer you paid, filmed on a phone.
Free stock video works if its licence allows edits. Pexels' licence, read on 10 November 2026, says "You can modify the photos and videos" and attribution is not required, but identifiable people may not appear "in a bad light" and you may not "imply endorsement".
A trend video from another creator is the grey zone. The steps of a trend are copied by millions of people; a person's actual video file is theirs. We used a public trend clip only for an internal test and did not post it. If you copy a trend, name the original creator in your caption.
Music: add the trending sound inside TikTok or Instagram instead of baking it into the file. Our guide on trending music for AI influencers has each platform's rules.
How do you make an AI influencer dance? Give a motion-control video model (for example Kling 2.6 Pro) a real dance clip to copy plus one picture of your character in the clip's opening pose. Prompting the dance in words doesn't hold the moves.
How much does an AI dance video cost? On fal on 10 November 2026, Kling motion control cost $0.07 to $0.168 per second. Our 13.8-second dance on 2.6 Pro came to about $1.55; on 3 Pro, about $2.32.
Is Kling 3 better than Kling 2.6 for dance videos? In our test, not by enough: the same moves and the same face, slightly sharper hands, for 50% more per second.
Why does my AI character's face change during the dance? Most often because the dancer in the reference clip turns away from the camera. Pick a clip where the dancer faces the phone, and fix the face in the first frame before animating.
Can I use any TikTok dance as the reference? Using someone else's video is a grey zone. Your own recording or a stock clip whose licence allows edits is safer, and the music should be added inside the app.
Make yours
Influgen's AI team can now film a dance this way: the creative director picks a studied trend clip from the trend board, draws the opening frame in your character's look, and the clip is face-checked before you see it. Create your own influencerfree here: 7-day trial with 50 credits, no card needed. Want to plan other shots first? The free AI video prompt builder writes shot lists, and Influgen in numbers has what our videos really cost.
Method
Two motion-control renders on 7 and 8 November 2026 of the same 13.8-second one-person dance clip and the same drawn first frame: fal-ai/kling-video/v2.6/pro/motion-control and fal-ai/kling-video/v3/pro/motion-control, orientation "video". Costs are fal's per-second price times the length. One earlier 10-second dance from a Pexels stock clip on 2.6 Pro, 4 November 2026.
Face scores are from our own face-matching model (cosine similarity against the character's reference photos).
Prices, input limits and the orientation description were read on fal's model pages on 10 November 2026; the Pexels licence on its licence page the same day.
Two clips and one character: a small test, not a benchmark.
Influgen gives your influencer a team that plans, creates and posts on its own. Create your own influencer free. The first 1,000 accounts get 100 extra credits.
Faces get all the attention, but clothes drift too: a coat that shortens, a fan that opens, a ponytail that comes down. Our AI critic rejected 15 of 59 frames for this in one week. Here are the 7 rules that came out of those misses.
Week two of our AI critic's log, 2 to 8 November 2026: 198 frame reviews, 69% kept (46% the week before). The glossy 'film still' look almost disappeared. The new top fault: pictures that give the punchline away before the end.
Before an Influgen influencer posts, a critic reviews every picture. From 26 October to 1 November 2026 it reviewed 274 frames and kept 46%. Here's what it rejected, in its own words, and the seven checks you can run on your own AI influencer content.
