Seedance 2.0 Prompts: Audio-Reference Test Checklist
Test a Seedance 2.0 audio reference with one limited role, a simple visual brief, and a full-output review—without assuming exact sound or sync.
For Seedance 2.0 prompts that use an audio reference, give one rights-cleared clip one narrow job—such as a timing cue or atmosphere reference—then test it with a simple visual brief and review the full result. Do not assume the file will be reproduced unchanged or that it guarantees speech, music, lip-sync, or a particular edit. A small test makes it possible to decide whether the audio role helped before you add more assets or publish anything.
Published September 29, 2026 · Audio-reference prompt checklist
ByteDance describes Seedance 2.0 as accepting text, image, audio, and video inputs in its official launch note. That is useful context, but the available controls depend on the service where you generate. The current Seedance 2.0 option in this studio allows up to three audio references and short 4–15-second outputs. This guide is deliberately narrower than a general prompt library: it is a method for testing one authorized audio asset without treating a generated clip as proof of an exact sound, spoken statement, or real event.
Decide what the audio is allowed to influence
Before uploading anything, write one sentence that states the audio reference’s role. A role is not a wish such as “make it cinematic.” It is an observation a reviewer can look for without assuming exact reproduction.
| Limited audio role | A useful test question | Avoid assuming |
|---|---|---|
| Timing cue | Does the visual action stay restrained around the intended beat? | An exact beat-for-beat copy |
| Atmosphere reference | Does the clip keep a quiet, compatible mood? | A specific licensed sound recording in the result |
| Pace reference | Does one simple visual change feel appropriately paced? | A fixed duration or edit rhythm |
| Review baseline | Can the team compare one version with and without the reference? | That the reference caused every difference |
Use a file only when you have the right to upload and use it for the planned purpose. A recording can contain a voice, music, location detail, or contract restriction that is not obvious from its filename. If a right, contributor, or permitted distribution scope is unclear, stop and use a different asset or get an answer from the person who controls it. The reference-image rights and quality checklist is also useful when an audio test includes a client image or recognizable person.
Keep the first visual brief quiet
An audio-reference test becomes unreadable when the prompt also asks for complex dialogue, a fast camera move, several scene changes, exact text, and multiple unrelated actions. Use one object or quiet location, one visible change, and a stable ending. The visual test should make it easy to spot drift.
For example, use a permitted ambient recording only as a timing reference for a simple still-life test:
A plain ceramic cup on a dark wooden table in soft window light. Steam rises slowly while the camera makes one gentle push-in. Use the cleared audio reference only as a quiet pacing cue for the scene. Keep the cup silhouette, handle, table edge, and light direction stable. One continuous shot. End on a centered still frame. No readable text, logos, extra objects, people, or scene change.
The wording does not promise that the reference will appear unchanged. It makes the visual acceptance checks clear: cup shape, camera restraint, light direction, and a usable ending. For the broader subject-action-camera structure, start with the controlled Seedance 2.0 prompt method.
Put only one audio asset in the first run
The studio may allow up to three audio references, but that is a ceiling rather than a recommendation. Begin with one. Name it in your project notes, state its narrow role, and keep the rest of the brief stable. A second or third file can make a later test harder to explain: a result may differ because of the asset, the combination, the prompt wording, or ordinary variation.
Use this pre-render check:
- Confirm the file is yours, licensed, or approved for the intended use.
- Write the audio role in one plain sentence.
- Select one simple visual subject and one visible action.
- Choose one quiet camera behavior.
- Name two or three visual details that must remain stable.
- State a final frame or hold that an editor can review.
- Record the date, model option, source files, prompt version, and the one question the test is meant to answer.
One reference with one stated role is easier to assess than a stack of files with overlapping instructions.
Review sound-sensitive work without overclaiming
Watch the whole result at normal speed, then inspect the opening, the busiest moment, and the ending. Your notes should describe what is visible and audible in the output, not assume a hidden cause. For example: “the camera remained gentle; the cup changed shape near the ending” is useful. “The reference forced the cup to change” is speculation.
| What you observe | Safe next decision |
|---|---|
| The visual brief holds and the pacing is usable | Keep the version as a limited concept and record the source and restriction notes |
| The visual brief holds but the intended pacing is not useful | Keep visual variables stable; revise one audio-role or timing instruction |
| A key object, person, or scene detail drifts | Reject the result for that use; simplify the visual brief or use verified media |
| Speech, music, a voice, or sync would need to be exact | Do not infer it from the render; use approved conventional audio and editing instead |
| You cannot verify the source or distribution permission | Stop the workflow until the input is cleared |
Do not publish a generated clip as if it documented a performance, voiced statement, event, or endorsement. The U.S. Copyright Office AI initiative is a source of current policy materials, not legal advice or permission for a particular voice, recording, likeness, or brand. For commercial work, check the actual agreement, platform terms, and applicable requirements before release.
Change one variable on the next version
If the first test is weak, do not replace the audio, prompt, subject, motion, camera, and ending at once. Keep a basic record and change the smallest thing that addresses the observation.
| Test note | Next controlled change |
|---|---|
| “The scene is visually stable, but the pace does not fit.” | Rephrase only the timing role or simplify the requested action |
| “The source and output length do not make the test useful.” | Use a shorter, clearer permitted reference or a simpler scene |
| “The cup changes shape during motion.” | Lock the camera or reduce the visual action; leave the audio role alone |
| “The output might be mistaken for a real claim.” | Reject it for publication and use verified footage or conventional editing |
This record protects the next reviewer too. Save the prompt, source identifiers, version, full output, observation, and decision together. The prompt test-log guide gives a compact way to compare revisions without inventing a general score for a model.
Frequently asked questions
How many audio references can I add to Seedance 2.0 on this site?
The current Seedance 2.0 selector allows up to three audio references. Use fewer when one cleared reference can answer the test question; adding more files does not make their roles clearer.
Can an audio reference guarantee exact speech, music, or lip-sync?
No. Treat a reference as a bounded input for a test, not a promise that audio will be copied unchanged or that speech, music, timing, or visual sync will be exact. Review the complete result before any use.
What should I change after an audio-reference test fails?
Keep the permitted source, visual brief, and review target stable, then change one variable: the audio role, the timing cue, the source quality, or the visual motion. Record the observed difference rather than guessing why it happened.
Make the next test easier to explain
Start with one cleared audio reference, one quiet visual job, and one review question. Keep only the output whose full clip supports that limited purpose; replace anything that needs exact sound, factual proof, or an unverified claim with approved conventional media. When the notes are ready, open the Seedance studio and run the smallest useful test first.