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 rather than a shipped product, with a public launch targeted for early July 2026; it slipped, and Seedance 2.5 was released on July 31, 2026, roughly five weeks after the reveal. 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.