ByteDance's Seedance 2.5 generates 30-second videos with 50 reference inputs

ByteDance has launched Seedance 2.5, an upgraded AI video generation model that pushes the boundaries of length, input variety, and editing precision. The model can generate native 30-second videos from a single text prompt—a significant leap from typical short clips. More notably, it accepts up to 50 multimodal reference inputs, combining text descriptions, static images, and video clips to guide the output. This allows creators to inject specific characters, styles, or scenes into the generation pipeline without retraining or complex workflows.

Seedance 2.5 also introduces local re-draw editing, a feature that enables users to modify individual elements within a video scene—such as an object’s color or a character’s expression—without regenerating the entire clip. This non-destructive editing capability reduces iteration time and compute costs, making it practical for professional video production, advertising, and content creation. By handling long-form video with rich context and surgical edits, Seedance 2.5 positions itself as a serious tool for creators who need both scale and control.