ByteDance presented Seedance 2.5 at the Volcengine FORCE conference, announcing a new generation of its text-to-video model with expanded capabilities.

Core technical improvements

The company highlights that the model can generate continuous video segments of 30 seconds in a single pass while preserving frame-to-frame coherence and scene logic.

• Native 4K output, produced directly by the model rather than by external upscaling networks.
• Support for up to 50 refs per textual prompt, including images, audio, and additional video files for fine-grained control.
• In-video inpainting, allowing users to replace specific details in a sequence without regenerating the entire clip.
• Control via 3D scene drafts that provide camera and character motion guides for texture-aware rendering.

For comparison, Seedance 2.0 accepted only 12 files as references, and ByteDance says the new limits enable more precise adjustments to character appearance, lighting, and soundtrack.

Commercial platform and rollout

ByteDance also announced a rights commercialization platform allowing copyright holders to license films and universes for AI-generated remixes within Douyin and Jimeng (Dreamina).

Revenue from views on generated works will be automatically split between content creators and studios, according to the company, although detailed mechanics were not disclosed.

The first official partner is Hong Kong director and actor Stephen Chow, whose comedy templates launched with the platform and produced over 100 000 user-generated videos within the initial twenty-four hours.

Availability and samples

Seedance 2.5 is currently in closed corporate beta for large enterprises, with a public release scheduled for early July. An official press release detailing technical specifications, sample assets, and pricing has not yet been published.