ByteDance's Seedance 2.5 stretches AI video to 30-second one-take clips

ByteDance has released Seedance 2.5, an update to its audio-video generation model that doubles single-pass output from 15 to 30 seconds and adds multi-round extension so creators can chain clips into coherent multi-minute videos. The pitch is a shift from generating isolated shots to producing finished narrative work: within a single generation the model is meant to sequence logically connected shots—setup, development, turning point, resolution—while holding characters, environments, and audio-visual pacing consistent across cuts.

The reference system is the other headline change. A single generation can now ingest up to 30 images, 10 video clips, and 10 audio clips as reference material, alongside capabilities like clay-render, motion, and creative referencing aimed at complex scenes with multiple subjects and shot changes. On the editing side, ByteDance adds timestamp-level control over audio and video, plus green screen, camera-perspective, and reference-based editing pitched at professional film and advertising workflows. The company also claims work on reducing the artificial look of AI video—textures, skin and eyes, lighting, saturation—and on suppressing stray subtitles and background music.

Seedance 2.5 is rolling out now through ByteDance’s own consumer apps, Jimeng AI and Doubao Pro, with API access promised soon via BytePlus ModelArk. The release continues the fast escalation of clip length, controllability, and multimodal conditioning among the major generative-video labs, and its explicit targeting of film and ad production signals ByteDance’s intent to compete for professional creative pipelines rather than just casual social clips.