中文

Why DanceDouble isn't “AI choreography”.

2026-11-20 · DanceDouble

“Import a song, get a dance” sounds a lot like the generative tools of the last few years, so people reasonably ask: is this AI choreography?

No. There is no neural network, no generative model, and no step that guesses at a picture. It does three plainer things: cut the song into beats, search a library of real movements for a sequence that lands on those beats, and render it on the device. Here is each of the three.

Step one: the song becomes a beat sheet

After you import an mp3, m4a or wav, the first thing that happens is not imagining a dance. It is spectral analysis: energy change per frame becomes an onset envelope — in plain terms, the moments where the music “moves”.

Then the app looks for a period in that envelope: it tries every tempo from 40 to 180 BPM, and at each tempo slides the phase (where beat one falls), looking for the alignment that absorbs the peaks most cleanly. The best-scoring one is the tempo, and it comes with a confidence figure — on a song with a vague beat, that number honestly comes out low.

Above that come downbeats and sections: which beats start a bar, which stretches look like a verse and which like a chorus. By now the song is a table: one row per bar, with its start and end time, its energy, and the kind of section it belongs to.

Step two: where the movements come from

The library holds 36 clips across three styles (groove, robot, hip-hop), 12 of each.

Eighteen of them are recorded motion capture: joint data from real dancers, chosen from the public motion capture database of Carnegie Mellon's Graphics Lab, retargeted onto the app's own skeleton and compressed to 25 keyframes each. CMU explicitly permits this data for research and commercial projects, so there is a person behind those moves.

The other eighteen are generated procedurally — hip sways, steps, arm circles, turns, built parametrically to fill the rhythmic patterns the capture set does not cover. In the library they are the same kind of object as the captured clips: each has a name, an energy value, the sections it suits, and a dominant side.

Step three: choreography is a search, not a generation

With a beat sheet and a library, the remaining question is: what goes in bar 1, bar 2, all the way to the end.

DanceDouble treats it as a search problem. Each bar is a slot, and each slot takes one of the 12 clips in the current style (and a left or right variant), so the whole song is one long sequence of choices. Finding the lowest-cost path through it uses a beam search: at each step it keeps the 16 most promising partial routes, extends and prunes them, and walks to the end of the song.

The cost is a sum of terms, each with a clear meaning and each tunable:

Together those terms make the search avoid the flaws you would spot instantly — a timid chorus, one clip looping four times, a landing pose that cannot lead into the next start. The whole thing is arithmetic, offline, and the same song with the same random seed always produces the same dance: rerun it, or change the seed to see a different arrangement, and the result is reproducible either way.

Step four: it dances for you on the device

Once the choreography is fixed, the skeleton goes onto a 3D stage: neon tubes, a reflective floor, a camera that pushes and pulls with the beat. The preview runs the same rendering path as the export, so what you see in the preview is what lands in the file.

Export is MP4 in vertical, square or widescreen framing, at 720P or 1080P, ready to post. Tempo analysis, choreography search and rendering all happen on the iPhone; the song is never uploaded anywhere.

So why not call it “AI choreography”

“AI” is used so broadly now that it nearly means “computed by a machine”. By that reading, a search is an algorithm too. But in most people's understanding, AI choreography or AI generation means: give a model a prompt or a song and it grows a sequence of movements that never existed.

DanceDouble is not that. Every movement in it comes from real data — either a person dancing or a parametrically built unit written by a person — and what it does is pick a run of them, under rules you can explain one by one. No frame is conjured, and no frame is “invented” by a model.

The distinction is not just wording. In mainland China, generative AI services go through a filing process under the Interim Measures for the Management of Generative AI Services, and DanceDouble's pipeline contains no generative model at all, so it should not and need not be declared as an AIGC product. “Search plus render” is both the most accurate description and the simplest one to stand behind.

There is a practical upside as well: a searched choreography can be explained, reproduced and tuned. Think the chorus is too repetitive — raise the repetition penalty. Think it is too frantic — lower the weight on energy matching. That is more grounding than rerolling a black box until something looks right.

See DanceDouble

App Store: id6835664114 · iPhone, iOS 18 or later · Free download, Pro is a one-time purchase
Not on the China mainland store yet (ICP filing in review); available everywhere else.

Keep reading

Products: Sealion Seacat Seaflower Seaspider Seasnipe Seastar Seadog Jiong WBXL Xultray webapp
Blink templateJS skiafy tranid Yimaji WenYaJi DanceDouble Pure Pace mkok i.zhuatang 原创歌曲
(C) 2026 ZHUATANG.COM, All rights reserved
This site counts visits with Google Analytics and a small first-party log (IP address, page and referrer, no cookies), only to improve these pages; the apps themselves ship no analytics SDK whatsoever.
媒体素材 / Press Kit · Blog

update: 2026-11-20