Seedance 2.5 advertises 30 seconds per generation and multi-round extension—but extension fails when you treat it like a longer text-to-video prompt. Editing fails when you do not tell the model what must survive. This guide covers both modes with ByteDance’s published grammar in mind.

Hard limits to plan around

30 seconds max per single pass (up from 15s on Seedance 2.0 family)
Up to 50 reference inputs (30 images + 10 videos + 10 audio) with documented duration caps
Native audio in the same pass—dialogue, SFX, ambience—not a separate render

Extension is sequential storytelling: each pass continues from a boundary frame, not a fresh prompt on a blank timeline.

Extension workflow

Generate a 10–15s master with locked subject language and camera height
Export the last frame (or use platform extend-from-clip)
Prompt only new action after the boundary—do not restate the entire first act
Audio: describe continuous ambience if you need seamless sound; hard cuts need explicit silence then new score language

Bad extend prompt: re-describe the character, location, lens, and wardrobe from scratch Good extend prompt: Continue from boundary: subject turns toward camera, same rain and coat, walks three steps, camera static.

Try Seedance in MidassAI

Video editing mode (preserve vs change)

Official guidance: name three buckets—

Master video (what the edit attaches to)
Edit scope (which seconds or regions change)
Content to preserve (face, logo, grade, camera angle)

Example: Master: clip A. Change only 00:08–00:12: replace product in hand with red can. Preserve face, background signage, and lighting direction.

Without a preserve list, Seedance treats the edit as a soft re-generation and identity drifts.

@ binding on extensions

When adding references on an extend job:

@Video1 = motion timing for arm gesture only. @Image1 = jacket color reference.

Do not attach ten references “for quality”—bind each asset to a single role or the model averages them.

Bracket syntax reminder

<> for SFX
{} for dialogue
【】 for subtitles

On extend passes, repeat only brackets that continue—new dialogue belongs in the extension segment, not duplicated from the master unless characters keep talking through the cut.

First / last frame tricks

For controlled endings, some workflows generate with an explicit last frame target (product hero pose). Align first and last before filling middle motion—reduces wobble on logo lockups.

Troubleshooting

Problem	Likely cause	Fix
Jump cut on extend	Boundary frame not aligned	Re-export last frame; reduce camera motion in master
Face morph mid-edit	Preserve list missing	Add explicit `preserve face geometry`
Audio restart	New score implied in prompt	Call out `continue ambient bed`

Cost discipline

720p 30s clips are materially more expensive than 5s probes on API meters. Storyboard at 480p / 4–6s, extend only on approved beats, then raise resolution for the final composite.

Platform note

Dreamina, Jimeng, and third-party APIs expose slightly different UI labels—verify Extend vs Edit in your host before burning credits. The grammar above maps to ByteDance’s Seedance 2.5 docs regardless of skin.

Seedance 2.5 Video Extension and Editing: A Practical Guide