Understanding the Core Signals: Motion, Pose, and Identity
Directing digital choreography requires separating three input layers: driving motion video, pose control signals, and identity reference images.
Motion transfer describes using a source performance to guide another subject. The implementation may use poses or other representations, so it is not the opposite of pose control.
Pose control describes guidance using body configurations, such as joint locations. A mask identifies an image region and is a different kind of signal. A system may combine these inputs, and guidance is not a guarantee that every output pixel obeys it.
Identity references are static still images (such as front-facing portraits) supplying facial likeness, skin tones, hair, and clothing textures.
In multi-person choreography, academic research such as the DanceTogether paper demonstrates how architectures separate identity references from spatial pose-mask streams to coordinate multiple dancers. While this highlights how experimental models isolate spatial guidance from character identity, it does not represent NEON DUO's production system.
Commercial systems use distinct proprietary approaches. For example, Runway Act-Two documentation combines a driving performance with a character reference for expressive motion. That tooling is model-specific to Runway's single-character setup; it does not indicate multi-person choreography support or reflect features on this site.
A generative workflow creates new visual content, while a timeline template may arrange existing media. Both categories contain different implementations. For an overview of how generative workflows differ from timeline templates, see our AI dance vs CapCut guide.
Which Dance Elements Can Be Guided (and Which Drift)
Not all dance components respond equally to control signals. Distinguishing controllable dynamics from drift-prone details sets realistic expectations.
Treat these as separate questions to verify in a product's documentation and actual output:
- Timing: Does the system accept a driving performance, and does the result follow the intended stops and accents?
- Body placement: Is pose input supported? Check arm and leg positions across the full clip rather than assuming the guide is followed exactly.
- Stage paths: Does each performer follow the intended route? Screen position and identity are different checks.
- Small details: Inspect hands, feet and faces at the intended viewing size. Input representations differ; do not assume every pose detector includes detailed fingers or that every generator uses a detector.
- Appearance: Compare clothes, hair and facial likeness separately from motion. Correct choreography does not establish stable identity.
A public paper or demo can illustrate a method. It cannot establish reliability for another service, another input pair or a different routine.
Diagnostic Decision Framework for Dance Choreography
Use this sequential checklist to determine whether a routine suits pose-guided generative workflows:
- Evaluate joint visibility: Verify that shoulders, elbows, hips, and knees remain visible throughout the source video without heavy shadows or obscuring clothing.
- Inspect occlusion frequency: Favor routines where dancers maintain open spacing over routines with frequent body overlap or tight embraces.
- Assess gesture dependency: Avoid routines reliant on intricate finger-tutting or hand signing, focusing instead on whole-body motion.
- Verify reference portraits: Confirm you have two separate, high-clarity adult portraits with clear lighting to anchor lead and support identities.
Worked Hypothetical Example: Directing a Two-Person Partner Turn
This illustrative worked example demonstrates how guidance signals interact during a partner underarm turn:
- Scenario parameters (illustrative): A 15-second Latin routine at an illustrative tempo of 120 beats per minute in 9:16 vertical orientation.
- Assigned roles: Partner A is the Lead; Partner B is the Support. Each role links to an individual adult portrait (under 10 MB in JPG, PNG, or WebP format).
- Opening: The planned shot shows both dancers holding hands. Record the intended identities and positions; do not assume they will remain stable in an untested result.
- Turn: The planned action hides one face temporarily. In a future authorized result, inspect that person before the turn and when the face becomes visible again.
- Evaluation: A change of sides is compatible with the choreography. A changed face is a separate issue. Check hands and role mapping independently.
- Decision: Accept or reject the actual output based on these observations. This hypothetical plan contains no measured outcome or claim that a model recovered successfully.
Frequently asked questions
Does supplying a clean pose skeleton guarantee crisp hands and fingers?
No. The input representation and output model both matter, and some pose representations contain little hand detail. Inspect the generated fingers; clean input alone does not establish accurate output.
Can pose control prevent two dancers from merging when they cross paths?
Not completely. While spatial masks help separate figures, deep overlaps—such as one dancer crossing behind another—challenge generative synthesis. You may observe momentary texture blending or swapped garment edges until separation returns.
Why does clothing style shift even when using the same choreography?
Scene styles share the exact same bound choreography, but visual appearance derives from identity portraits and style prompts. If reference photos feature ambiguous textures or loose garments, secondary clothing motion may vary across generated clips.
Can I generate a two-person dance from a single group photograph?
NEON DUO requires two separate image files of consenting adults, in JPG, PNG or WebP, up to 10 MB each. That is this interface's requirement, not a rule for every multi-person model. If you crop from a shared photograph, check permission, clarity and framing for each person.
Next Steps: Preparing Assets for Guided Motion
Directing AI dance sequences requires choosing choreography suited to spatial guidance while avoiding tight occlusions and intricate hand gestures.
While live AI generation remains temporarily closed on NEON DUO, you can explore interface layouts and sample references on our create workspace or browse technical breakdowns in our guides index.
To prepare compatible assets:
- Select two distinct, well-lit adult reference portraits with clear, unobstructed facial features.
- Save your source files as standard JPG, PNG, or WebP images under 10 MB each.
- Review our published terms of service and privacy policy to understand data usage and asset handling guidelines.