Seedance 2.5 doubles one-pass video length and adds timestamp-level editing

ByteDance's Seed team has released Seedance 2.5, an audio-video joint-generation model that can produce a 30-second clip in one pass. That doubles the single-pass length the team states for Seedance 2.0, from 15 seconds to 30. Users can extend a clip through multiple rounds, but that is different from generating one uninterrupted several-minute video. 1

The release is aimed at longer, more coherent sequences. ByteDance says it improved shot transitions, scene changes, audiovisual continuity, and the stability of subjects across cuts. The official announcement uses qualitative language rather than publishing benchmark scores, so those quality claims are not yet independently comparable. 1

A single generation can take up to 30 images, 10 video clips, and 10 audio clips as references. That lets a creator specify several characters, scenes, camera changes, or sound cues instead of building the sequence from one reference image at a time. The model also accepts prompts for narrative, camera perspective, movement, and rhythm. After generation, timestamp-level editing can target characters, actions, plot elements, backgrounds, and camera perspective. 1

That combination changes the practical unit of work: a creator can start from a small audiovisual brief, then revise a specific moment rather than regenerate the whole clip. It is a workflow claim, not proof that every difficult scene will hold together.

Seedance 2.5 is rolling out on Jimeng AI, Doubao Pro, and other platforms. ByteDance says API access is coming through BytePlus ModelArk, so developers who need programmatic testing still have to wait for that route to become available. 1

Independent coverage confirms the 30-second audio-video workflow and notes that Google's Gemini Omni Flash produces roughly 10-second clips, giving Seedance a longer single-pass target. It does not provide an independent Seedance 2.5 benchmark, either. 2 A Hacker News submission linking the announcement had 354 points and 187 comments when retrieved, evidence of developer attention rather than a quality evaluation. 3

ByteDance explicitly says complex-motion plausibility and scenes involving interactions among multiple subjects still need work. For now, the release is best understood as a more capable video-production workflow with wider reference and editing controls—not as a measured across-the-board lead. 1