Write Seedance 2.5 timestamp prompts using the time inside the clip you are generating or editing, then place that clip against the full song in your editor. Keep a separate note of where the clip belongs in the recording. Mixing those two clocks can turn a clear creative idea into an ambiguous request.
ByteDance's Seedance 2.5 announcement describes timestamp-based generation guidance and targeted audio/video editing. It also describes clips of up to 30 seconds. These capabilities do not establish frame-perfect obedience, and the available duration and editing controls still depend on the interface you use.
This guide develops an original 12-second sequence and a repair request. The examples are editorial planning exercises, not tested presets or measured results.
Translate the song clock into the clip clock
Imagine that a visual passage should run from 1:12 to 1:24 in your final song. The clip is 12 seconds long, but its own timeline begins at zero.
Use this relationship when planning:
Clip time = song time minus the clip's placement time
Song time = clip placement time plus clip time
For this example:
| Event | Position in the song | Position inside the clip |
|---|---|---|
| Passage starts | 1:12 | 0 seconds |
| A lid begins to rise | 1:16 | 4 seconds |
| The contents become visible | 1:20 | 8 seconds |
| Passage ends | 1:24 | 12 seconds |
The timestamps are invented for illustration. Mark the actual moments in your selected recording before writing the prompt. A tempo estimate is not enough if the singer pauses, stretches a phrase, or enters before the bar.
Keep a placement note outside the generation prompt: Clip A belongs at song 1:12. Inside the prompt, describe a 0–12-second sequence. Do not expect a model receiving a short clip request to infer where 1:12 falls in a separate, unattached song.
Separate the scene description from the event schedule
First describe what stays true for the whole sequence. Then describe what changes. This makes it easier to spot contradictions such as asking for a stationary camera at the top and a circling camera halfway through.
Our scene uses a small blue box on a worktable. Its lid opens to reveal a warm light. There is no performer, readable text, or complicated hand interaction to solve alongside the timing question.
Set a supported 12-second duration in your interface, if available, and choose the intended frame shape there. The following text communicates the visual plan; it does not replace those controls:
One continuous view of a small matte blue box on a pale wooden worktable.
An uncluttered workshop background remains softly out of focus.
The camera is stationary at table height. Keep the whole box visible.
Cool daylight lights the room throughout the sequence.
0–4 seconds: the box stays closed. A faint warm line is visible beneath the lid.
4–8 seconds: the lid rises gradually on its rear hinge. Warm light becomes
visible inside the box as the opening grows.
8–12 seconds: the lid stays open. The warm light settles to a steady glow,
with the box and the surrounding table remaining still.
The box keeps the same size, color, hinge position, and place on the table.
No writing or symbols appear on the box.
If your provider does not support the duration or operation you need, adapt the sequence to its documented controls. Do not rename a different model as 2.5 or assume a prompt can unlock a missing feature.
For a broader capability comparison, see Seedance 2.0 versus 2.5. This article is about organizing a timed request, not choosing a universal winner between versions.
Give each time window a visible outcome
“Build anticipation” describes an intention. “A faint line of light appears beneath the closed lid” gives you something to inspect. Tie the emotional purpose to an observable change.
In the example, the opening is the principal event. The first window establishes the object, and the last provides a stable ending. If you also request sparks, flying tools, a camera orbit, and a room transformation, it becomes harder to identify why the lid opens late or the box changes shape.
Before generating, read the schedule for physical continuity:
- Does the next action start from the state left by the previous one?
- Is the same object still in the same place?
- Is the camera behavior consistent?
- Have you left time for the main action to finish?
Use a few meaningful windows rather than a separate instruction for every tiny movement. The purpose of the schedule is to communicate the sequence, not to create an animation keyframe file that the model has promised to execute exactly.
Review the returned action, not just the requested timestamps
Put the result on the song timeline and check where the lid actually starts moving and where the reveal becomes clear. Record those observed times separately from the planned ones.
Suppose the requested reveal was at clip second 8, but the useful reveal arrives at second 9. Your planned song position was 1:20; the observed event now lands at 1:21 when the clip starts at 1:12.
That difference gives you a concrete editing question. Moving the clip one second earlier would align the reveal, but it would also move the opening and ending. Check whether the preceding shot and the next cut still work before making that adjustment.
| What happened | First thing to try |
|---|---|
| The event is good but slightly late | Test placement or trimming in the timeline |
| The lid barely starts before the clip ends | Simplify the action or allocate more of the supported duration to it |
| The box changes into another object | Reduce competing visual instructions and review references |
| The reveal aligns but the opening no longer fits | Reconsider the preceding cut rather than forcing the whole sequence to match one point |
| Generated sound conflicts with the song | Review the clip's audio controls or mute unwanted clip audio during editing |
These are editorial diagnoses, not guaranteed model fixes. Keep the original result while evaluating a revision.
Name the time range of the actual edit input
A repair request needs its own clock check. If you upload the entire 12-second output, second 6 means six seconds from that file's beginning. If you first trim it to only the opening action, the same visual moment has a different timestamp.
For example, exporting only original clip seconds 4–10 creates a six-second file. Original second 6 is now second 2. Verify the uploaded file before writing a targeted edit request.
An illustrative repair for the untrimmed 12-second clip could be:
In the supplied video, revise the lid motion between seconds 4 and 8.
The lid should rise on its existing rear hinge instead of bending at the center.
Preserve the blue box, its position, the stationary camera, and the workshop.
Keep the closed opening state and the open ending state consistent with the source.
Attach the source using the editing controls your interface supports. A reference to “the supplied video” does nothing if no video is attached. ByteDance documents timestamp-directed edits, but preservation instructions remain requests that you must check in the returned result. Official editing description.
Watch the entire edited clip, including the entry and exit of the changed range. A repaired hinge is not an improvement if the object jumps at second 4 or the final frame no longer fits the next shot.
Keep timing notes with the approved audio
Save three items together: the final song file, the accepted clip, and a short placement note. Include any trim you applied. This prevents an editor—or your future self—from having to reconstruct which clock a timestamp belongs to.
If you change the song version, recheck placement against the new recording. If you change playback speed, the simple offset table above no longer describes the timing without accounting for that speed change.
For assembling the complete song, use our music-video production workflow. LyricVideoMaker provides an audio-upload and editable lyric-video workflow; this external Seedance prompting method does not imply a built-in Seedance 2.5 editing integration.
Are prompt timestamps enough to synchronize singing?
No. A timed visual instruction does not establish that mouth movements match the words in your recording. If the scene includes a singing face, use the checks in the music-video lip-sync guide.
Should every clip use the model's maximum duration?
No. Request the supported length that fits the visual task and your edit. A short, clearly defined event can be easier to review than a longer sequence containing several unrelated changes.
What should I change first when timing is wrong?
Confirm the source file and time origin, then inspect the actual event in the returned clip. Try an edit-timeline adjustment when the action itself is usable. Regenerate or repair when the action—not merely its placement—needs to change.