Seedance 2.5 by ByteDance turns a single prompt into a 30-second 4K video, with up to 50 references and native audio sync — great for music videos, ads, and cinematic clips.
2026.07.01
Seedance 2.5 is ByteDance's next-generation video model that turns a single prompt into a 30-second 4K video in one pass. Where most AI video models cap out around 15 seconds, Seedance 2.5 generates one continuous scene without stitching short clips together.
It accepts up to 50 multimodal references — images, video, audio, and style inputs (up from 12 in the previous generation) — so you can control the subject, tone, and composition far more precisely. Sound is generated in the same space as the visuals, so onscreen action and audio stay in sync, and ByteDance reports about 20% better prompt adherence, meaning fewer regenerations before a usable result.
Seedance 2.5 is currently in enterprise beta ahead of its public launch. It isn't available in Carat yet, but you can create 4K video today with Seedance 2.0, which uses the same video grammar.
Seedance 2.5's strength is generating longer video in one continuous take — with sound included. A full 30 seconds of setup, development, and payoff can be designed in a single prompt, cutting out the old workflow of making and stitching separate clips. Locking a subject, product, or mood with up to 50 references keeps the same face and tone throughout, and because motion and audio are generated together, effects like footsteps and ambient sound land in sync. For now, Seedance 2.0 is the closest model you can run in Carat with the same grammar (@Image1 references, Shot and timecode multishots, tracking and dolly moves), so practicing there is the best way to get ready for 2.5.
Shot 1: / Shot 2: or 0.0-2.5s timecodes.@Image1, @Image2, and so on — up to 50 inputs.Until Seedance 2.5 is supported, get familiar with the same grammar in Seedance 2.0. In Carat you can generate 4K video with Seedance 2.0, then keep going with follow-up work like changing the ratio, adding captions, and upscaling.