Seedance 2.5 is ByteDance's video generation model, built by the company behind TikTok: it renders a continuous 30-second shot in one pass, soundtrack included. The model was introduced at Volcano Engine's FORCE conference in Beijing, skipping versions 2.1 through 2.4 to mark the jump from Seedance 2.0. Access runs through Dreamina, the international interface tied to CapCut, on a credit system, or through a usage-billed API. Ad studios and video teams get the most out of it.
Each request accepts up to 50 references: 30 images, 10 video clips and 10 audio tracks. That locks a character, a set and a camera style before you even write the prompt. You drop three product photos and a clip for the camera move, and the scene comes back framed the way you planned. A single audio track can even set the pace, with cuts and lip-sync following the beat.
Output climbs to 4K, with dialogue and sound effects generated alongside the picture in over ten languages. When one detail goes wrong, region-level editing fixes a portion of the frame without rerunning the whole clip (a six-fingered hand no longer sends a good shot to the bin). A long format, still in beta, stretches a single generation to 180 seconds.
| What changes | Seedance 2.0 | Seedance 2.5 |
|---|---|---|
| Shot length | 15 seconds maximum | 30 seconds in one pass |
| Reference inputs | 9 images and 3 videos | 30 images, 10 videos, 10 audio |
| Fixing a result | full regeneration | region-level frame editing |
| Long format | none | up to 180 seconds in beta |
Seedance 2.0, its predecessor, took first place in the Artificial Analysis video arena at release, for text-to-video and image-to-video alike. Version 2.5 pushes where rival AI video generators hit a ceiling: Kling 3.0 assembles up to 15 seconds across multiple shots and Veo 3.1 chains scene extensions, while Seedance renders its 30 seconds as one block, with no seams between segments.
The model itself stays closed, with no weights to download, in contrast with Alibaba's open Wan line. ByteDance's Seed team does publish plenty of research in the open, yet keeps its house video model server-side. The upside lands on your desk: no GPU to buy, full quality straight from the browser or the API.
Seedance 2.5 runs on Dreamina internationally, on Jimeng in China and inside the Pro tier of the Doubao assistant. Developers call the API on Volcano Engine Ark or BytePlus, billed on usage with no free quota.
On Dreamina, paid plans remove the watermark and cover commercial use; a 30-second 4K shot logically draws more credits than a short clip. On the API side, third-party estimates place a 30-second render between $3.60 and $9. Those figures will shift, so treat ByteDance's live rate card as the only one that counts.
Not for heavy work. Dreamina is the simplest way in, with generations drawing on the credits of your subscription; paid plans add commercial licensing and drop the watermark. On the API side there is no free quota, so billing starts with the first call. Check the current rate card before committing to a large project.
No, Seedance 2.5 is proprietary: ByteDance has released neither weights nor inference code, and the model runs through Dreamina, Jimeng or the API (also available on Higgfields). The Seed team does share other research openly, yet the Seedance family stays server-side. One quiet benefit: nothing to install, and full 4K quality without dedicated hardware.
It depends on the format. Kling 3.0 has a free tier and a multi-shot mode around 15 seconds, solid for brief clips. Seedance 2.5 takes the lead once you want a continuous 30-second shot, a stable identity over time or a large stack of references. For sustained storytelling, it holds a clear edge.
The reference page sits on seed.bytedance.com, the portal of ByteDance's Seed team, while public access goes through Dreamina at dreamina.capcut.com. Be wary of independent domains that borrow the Seedance name in their address: they are not affiliated with ByteDance and mostly resell access, sometimes with shaky information.
Verdict: Product ads, music videos, short fiction: the continuous 30-second shot removes the cuts that give generated footage away. Brands, studios and social media teams after characters that hold steady from one shot to the next have their production engine here.
