The weirdest way to make AI copy a dance better? Take the dancer out of

提示词

The weirdest way to make AI copy a dance better?

Take the dancer out of the video.
Sounds backwards.
But this simple trick can dramatically change how motion transfer works in Seedance 2.0.

When you give a video of someone dancing directly to a video model, you’re not giving it “motion.”

You’re giving it everything:

The person.
Their face.
Their clothes.
The room.
The lighting.
The camera.
The background.
And somewhere inside all that noise — the choreography.

Now ask the model to put that movement onto a completely different character.
You’re essentially asking it to solve ten problems at once.
And then we wonder why the choreography drifts.
The trick is to separate two things that shouldn’t be mixed:

WHO is moving
and
HOW they are moving.

Here’s the workflow:

1. Lock the character first.
Create the character reference in GPT Image 2 — face, body, clothing, overall identity.
The goal is simple: the character’s identity should exist independently from the movement.

2. Strip the dancer out of the source video.
Convert the dance footage into a depth map.
Instead of giving the model the original performer, you give it spatial information — essentially a representation of where the body is in 3D space.

No face.
No clothes.
No background.
No visual identity.
Just movement.

3. Give the model two separate references.
Character sheet = identity
Depth map = motion
Now each input has one job.
And that’s the important part.

4. Don’t let the references compete.

The depth map shouldn’t tell the model who is dancing.
The character reference shouldn’t tell it how to dance.
You’re removing the ambiguity instead of asking the model to resolve it.
That’s why this works.

Raw footage contains an enormous amount of information, and the model has to decide what matters.

A depth map dramatically narrows the problem.

Instead of:

“What am I supposed to copy from this video?”

the model gets something closer to:

“Here is the movement. Put this character inside it.”

And that distinction matters.

Because better AI video isn’t always about giving the model more information.

Sometimes it’s about giving it less — but making every piece of information intentional.

This opens up some interesting workflows:

→ Transfer choreography onto completely original characters
→ Create motion-capture-style results without a mocap suit
→ Preserve specific movements more accurately
→ Build character showcases from existing performance footage
→ Separate identity, motion and environment instead of letting them bleed into each other

The character sheet answers:
“Who is dancing?”

The depth map answers:
“How are they moving?”

And keeping those two questions separate may be one of the most useful tricks for getting more reliable motion transfer out of AI video models.

The model doesn’t necessarily need to become smarter.

Sometimes you just need to stop giving it the wrong problem.