30 seconds of video, all at once, with no stitching clips together. This is what Seedance 2.5 does, made by ByteDance. It accepts up to 50 references (image, video, audio) so you can control the result. The official pages don't mention 4K: the plans list 480p, 720p and 1080p.
- ByteDance's Seedance 2.5 generates up to 30 seconds of video in a single pass, with no stitching of clips.
- It accepts up to 50 reference inputs (image, video, audio) for a consistent character, location, and style.
- Officially launched on 31 July; the official pages don't mention 4K, and the emphasis is on control.
We claimed Seedance 2.5 makes 30 seconds of video in native 4K, accepts 3D references, that an enterprise beta was live and that the public launch was in early July. The official Volcano Engine and ByteDance Seed pages confirm the 30 seconds and the 50 references, but don't mention 4K (they list 480p, 720p and 1080p), 3D references or an enterprise beta; the official launch came on 31 July. The headline, dek and text were corrected and the official sources added. The conclusion about control through references doesn't change.
ByteDance showed Seedance 2.5 - a model that generates a 30-second video in a single pass, without the usual stitching of short clips. More important for actual work is the control: it accepts up to 50 reference inputs at once - image, video, audio.
I've shot scenes where the whole evening rode on one thing not falling apart - light, set, face, everything staying the same from start to end. The designer's pain with AI video has always been exactly this: not 'will I get a good image,' but 'will my character fall apart in the next frame.' The references hit exactly there.
Until now, AI video was flashy and unmanageable. With 50 references you plan the shot in advance, instead of spinning generations and hoping. You're not guessing anymore. For production, that weighs more than the seconds themselves.