ByteDance unveils Seedance 2.5: 30-second clips in a single pass

The ByteDance Seed team has announced the release of Seedance 2.5 — the latest version of its video generation model. According to the developers, user requests have shifted from “make a clip” to “make a finished piece,” and the new model is built around two tasks: base generation and reference-based generation.

What's new

The main change is duration. Seedance 2.5 generates up to 30 seconds of audiovisual content in a single pass (versus 15 in the previous version) and supports repeated extension, letting you assemble clips several minutes long without manually stitching and adjusting transitions. The model, it is claimed, builds several logically connected shots within a clip — with a setup, development, twist and resolution — rather than simply stretching out a single moment.

The set of references has been expanded as well: in a single pass you can supply up to 30 images, 10 video fragments and 10 audio tracks. The model takes into account composition, scenes, styles, characters and props, preserving the appearance and voices of several protagonists even in group scenes. Also mentioned separately are clay render mode, motion references and creative references: from a texture-free 3D model you can set the spatial structure, poses, trajectories and camera angles, and the model will fill in physically plausible lighting.

Editing by timecode

Seedance 2.5 supports targeted editing tied to timestamps: you can control rhythm, angle and movement on a specific segment, and after generation edit characters, actions and plot details without breaking continuity. “Professional” features are also announced — green screen work (replacing the background while preserving the protagonist), angle changes and reference-based editing, aimed at film and advertising.

Beyond entertainment

The developers specifically highlight practical scenarios. In education, the model turns historical context and scientific principles into illustrative clips for lessons. In industry, robotics and autonomous transport, Seedance 2.5 generates synthetic video data for training perception systems and models rare scenarios — extreme weather and complex road situations.

At the same time, the team honestly acknowledges its weak points: the physical plausibility of complex movements and the stability of scenes involving several interacting characters still leave something to be desired. Seedance 2.5 is already rolling out in the Jimeng AI and Doubao Pro apps; API access via BytePlus ModelArk is promised later.

Source: Hacker News