Seedance 2.5

ByteDance has officially launched its next-generation video generation model, Seedance 2.5. Building on Seedance 2.0's integrated audio-video generation architecture, it strengthens long-form storytelling, multimodal reference, and video editing capabilities.

Key features are as follows.

• Single generation of up to 30 seconds: Generate high-quality 30-second videos with audio in a single pass, with the ability to extend multiple times. Improved scene transitions and shot connections support the creation of consistent content lasting several minutes.
• Multimodal reference input: A single generation can take up to 30 images, 10 videos, and 10 audio clips as reference material. It supports various reference types such as clay renders, motion, and creative references.
• Precise editing control: Timestamp-level editing is possible for both audio and video, with improved green screen, camera viewpoint, and reference-based editing features as well.

Rather than simply extending a single scene, the model aims for long-form narrative generation that composes multiple shots—including setup, development, transition, and ending—within 30 seconds. In the official example, a singer moves from the dressing room to the stage, interacts with dancers, and then begins the performance, all expressed as a single continuous video.

Seedance 2.5 is currently available on Jimeng AI and Doubao Pro, among others, with BytePlus ModelArk API access to be supported later.