Seedance 2.5 Is Live: How to Prompt 30-Second Single-Take Scenes

ByteDance first showed Seedance 2.5 at its Volcano Engine FORCE conference back in June, and Karya covered it then as a closed enterprise beta. That changed this week: the public developer API opened on August 7, 2026, with access first through BytePlus ModelArk and a rollout across ByteDance's Jimeng, Dreamina and Doubao apps. It is no longer a preview - creators outside the beta list can now actually generate with it.

What shipped

• Up to 30 seconds of video in a single generation pass, with synchronized audio generated alongside the visuals rather than added afterward.
• Multi-round extension - you can continue a clip past 30 seconds while the model keeps character, environment and pacing consistent, instead of stitching separate generations together in an editor.
• Up to 30 reference images, 10 reference video clips and 10 reference audio clips in one generation - a large jump from Seedance 2.0's 9 images, 3 videos and 3 audio clips.
• Timestamp-level editing, so you can revise one section of a clip without regenerating the whole thing.
• Better handling of multi-subject scenes and complex physical motion, according to ByteDance's release notes.

The real change is how you write the prompt

A 5-second clip rewards a single vivid description: one action, one camera move, one mood. Thirty seconds does not work that way - if you write it like a single moment stretched thin, the model has nothing to do with the extra time and the scene drags. Seedance 2.5's own prompt guidance leans into describing an actual arc: a setup, a development, a turning point, and a resolution, the same shape you would use to plan a short scene in a script.

The reference system changed shape too. Instead of dropping in one reference image and hoping the model infers what it is for, label each reference by role directly in the prompt - character, location, palette, product - so the model applies each one to the right part of the scene instead of guessing. With up to 50 references available across images, video and audio, an unlabeled pile of files is more likely to confuse the generation than help it.

How to get ready even before it lands in your tool of choice

• Build a labeled reference set now for any recurring character, product or location - a clean folder of "character front/side/back", "location wide/detail", "product hero shot" beats scrambling for references after the fact.
• Practice writing scene-shaped prompts instead of moment-shaped ones: block out an opening beat, a middle development and a closing beat before you write a single sentence of prompt.
• Keep a short written style anchor (lighting, color grade, camera language) you reuse across prompts, so a longer clip stays visually consistent from the first frame to the last.

Using Seedance today in Karya

Karya's Video Studio runs Seedance 2.0 today - Standard, Fast and Mini tiers, for both text-to-video and image-to-video, with multi-image reference support and the credit cost shown before you generate. The scene-first, multi-beat prompting habit above is worth building now: it already improves longer Seedance 2.0 generations, and it is exactly the skill Seedance 2.5 rewards once it is available inside Karya.