Seedance 2.5

Seedance 2.5, announced by ByteDance Seed at the end of July 2026, is the generation in which the company's video model stops producing clips and starts producing scenes. The generation ceiling doubled from fifteen to thirty seconds, and - the part that matters more than the number - those thirty seconds are one coherent take rather than segments stitched together, so a short story can be told without a cut imposed by the tool. Sound is generated with the picture in the same pass rather than laid over it afterwards, and a reference track of speech, music or effects can be the only input, with the model matching pacing and lip movement to it. The more unusual capability is what ByteDance calls omni reference: a single request may carry up to fifty assets - thirty images, ten video clips and ten audio files - which the model reads together to decide what the shot should contain. That opens three jobs beyond plain text-to-video. It can edit an existing video by timestamp, replacing a subject, removing an object or repainting part of the frame while keeping the original aspect ratio and length; it can extend a clip forwards or backwards; and it can generate from a first frame, or from a first and last frame, with the motion in between inferred. The output format was widened to mov with H.264 and PCM audio to keep colour and sound intact through editing. One correction to the figures circulating in the trade press: the model card on BytePlus ModelArk lists 480p and 720p as the output resolutions for Seedance 2.5, while the older Seedance 2.0 is the model listed up to 1080p and 4K. Reports of a 4K flagship appear to conflate the two. The model is closed - available through Jimeng, Doubao and the ModelArk API, with no published weights, parameter count or technical report.

▸News

ByteDance's new video model holds one shot for thirty seconds - and its API card quietly says 720p, not 4K

8/11/2026 · BytePlus ModelArk documentation

▸Videos

No videos yet.

▸Reviews