models

Wan vs Seedance Prompts for AI Rap Video

The same AI rap video render, run through a Wan-style pipeline and a Seedance-style pipeline, wants slightly different words. Not a different workflow — the templates on this wiki are deliberately portable — but a reordering: what goes first, how references are named, where the negative block lives. This page is the translation table plus a worked example in both dialects.

The one structural difference

Wan-style pipelines are subject-first. They reward prompts that define the cast before the set: "Left performer = photo 1, right performer = photo 2", roles and rules up top, stage second. Seedance-style pipelines are scene-first. They reward atmosphere before cast: the stage, the light, the mood — then who's standing in it, described by appearance. Both understand the same vocabulary; they just weight the opening of the prompt most heavily, so the most important instruction goes first in each dialect.

Dimension-by-dimension

Copy-ready: the same duo prompt in both dialects

Wan-style (subject-first)

Create an AI rap video duo performance from two separate portrait references.
Left performer = photo 1, right performer = photo 2. Fixed sides for the whole clip.
Rules: only one performer raps at a time; the other nods on beat. Faces stay sharp and centered.
Stage: original studio booth, warm orange key light, one hanging microphone centered between them, dark backdrop.
Action: energetic head nods, traded bars with clear mouth movement, one slow push-in during the final two seconds.
Audio: original upbeat rap cadence, call-and-response — write NEW lyrics, never quote commercial tracks.
Duration: 8-12 seconds, clean two-shot hold at the end.
Negative: identity swap, merged faces, third person, blurry faces, on-screen text, logos, NSFW.

Seedance-style (scene-first)

An original studio booth stage: warm orange key light falls from the front-left, a single vintage microphone hangs centered in frame, the dark backdrop carries a soft rim light.
Two performers stand on either side of the mic: on the left, the person from the first reference photo in their own outfit; on the right, the person from the second reference photo. They trade bars in a call-and-response — one raps while the other nods on beat — energetic but locked to their sides, faces sharp and front-lit.
The camera holds a medium two-shot, pushing in slowly over the final two seconds to a clean hold. The performance runs 8-12 seconds to an original upbeat rap cadence with NEW lyrics — never commercial songs.
Avoid: identity swap, merged faces, third person, blurry faces, on-screen text, logos, NSFW.

Notice what stayed identical: fixed sides, one-rapper-at-a-time, the mic anchor, the loop hold, the original-lyrics guardrail, the negative terms. The dialect is ordering and naming — the stage vocabulary ports verbatim.

Porting rules

1. Move the crown jewel to line one. Subject roles first for Wan-style; stage atmosphere first for Seedance-style.
2. Rename references. "photo 1" becomes "the person from the first reference photo" — appearance-based naming survives both.
3. Halve the choreography list for Seedance-style. Keep the two strongest gestures; fold the rest into "traded bars with clear mouth movement".
4. Relocate the negative block. Dedicated field if it exists; otherwise "Avoid:" as the final line — never the opening.
5. Re-render, don't assume. The first render in a new dialect is a calibration shot; expect one wording pass.

Common failure modes

• Scene-first prompt in a Wan-style tool drifts sides → the role definition buried too deep; hoist "Left = photo 1" back to the top.
• Subject-first prompt in a Seedance-style tool renders a flat stage → atmosphere arrived last and got clipped; open with the stage.
• Choreography half-applied → the list was long for the dialect; cut to two named gestures.

FAQ

Which style is better for AI rap videos?

Neither universally — identity-heavy renders (self-duo, pets) tend to favor the explicit slot naming of Wan-style; atmospheric stages (rooftop couple) favor scene-first. The wiki's templates are written neutral so they run in both with the porting rules above.

Do these rules apply to specific model versions?

They describe behavior patterns shared across pipeline families, not pinned versions. Check your tool's own docs for its negative-field and duration specifics — the porting rules absorb the rest.