Seedance 2.5 promises 30-second AI videos with detailed creative control

ByteDance has launched Seedance 2.5, a video generation model designed to create longer clips, use more reference material and support more precise edits. ByteDance announces the release in an official Seed blog post, describing the model as a step from generating isolated clips toward producing more complete audiovisual work.

The model can generate up to 30 seconds of video with audio in one pass, double the 15-second limit ByteDance cites for the previous version. Users can then extend a generated clip in several rounds. According to the company, this process is intended to preserve the appearance of key characters, the setting, pacing and sound across longer sequences.

That capability could matter for teams producing short social videos, advertising concepts, training materials or storyboards. Rather than assembling many unrelated clips, creators can attempt to build a sequence with connected scenes, camera movements and plot developments. ByteDance says Seedance 2.5 can organize a short narrative with an opening, development and conclusion within a 30-second generation.

More references for complex scenes

Seedance 2.5 accepts a larger set of source material in a single request. Users can provide up to 30 images, 10 video clips and 10 audio clips as references. The company says the model can draw on these materials for character appearance, voices, locations, props, composition, visual style and movement.

This approach is aimed at productions that require consistency across multiple people or settings. ByteDance presents examples involving concert scenes, opera performances and animated sequences, where individual reference images define performers, venues and supporting characters.

The update also expands what ByteDance calls clay render referencing. A creator can supply a simple, textureless 3D scene to specify camera positions, subject movement, spatial layout and blocking. Seedance 2.5 is then intended to render that basic structure as a finished video with materials, lighting and a chosen visual style. The company says the model uses the spatial information to make shadows, light direction and colour temperature appear more physically coherent.

Timestamp-based changes

The other central feature is more detailed editing. Seedance 2.5 supports instructions tied to timestamps, allowing users to specify what should happen during particular seconds of a clip. They can direct camera perspective, movement, plot elements and rhythm during generation, then request targeted changes after a clip has been created.

ByteDance also highlights green-screen editing, camera-motion editing and reference-based editing. In one example, users replace backgrounds and obstacles around a person while retaining the main subject. The company says the system can adapt clothing movement, hair, gait and lighting so that the person fits the replacement environment.

For content professionals, these functions may reduce the need to regenerate an entire clip after a small mistake or creative change. However, the company acknowledges that the model still has limitations. It identifies the physical plausibility of complex movement and the stability of scenes with several interacting subjects as areas that need improvement.

Seedance 2.5 is rolling out through ByteDance’s Jimeng AI and Doubao Pro services. The company says API access will follow through BytePlus ModelArk. ByteDance also positions the system for education, industrial simulation, robotics training and autonomous-driving data generation, alongside creative uses in film and advertising.