At the Volcano Engine FORCE conference in Beijing, ByteDance unveiled Seedance 2.5 — a model it claims can generate a single continuous 30-second clip in one coherent diffusion pass, with no stitching of shorter segments. A unified joint audio-video architecture co-processes visual and audio signals in the same latent space, holding character appearance, lighting, and motion style across the full half-minute. The bigger leap for creators is reference capacity: Seedance 2.5 accepts up to 50 simultaneous multimodal reference materials — images, video clips, and audio files — versus roughly 12 in Seedance 2.0, and against just three for Google's Veo 3.1. That headroom is built for brand-consistency and episodic storytelling, where a face, a product, an art style, and a music reference all have to stay locked across a generation. Three modes ship — text-to-video, image-to-video, and a motion-reference mode that copies camera and action style from an existing clip — alongside localized region editing that lets a creator redraw one element of a frame without regenerating the whole shot. Announced as a global enterprise beta, not a shipped product, with a public launch targeted for early July 2026 and no pricing disclosed; all capabilities are ByteDance's own claims, unverified by independent benchmarks. The launch lands with its predecessor's copyright disputes from every major Hollywood studio still unresolved.