ByteDance has introduced Seedance 2.5, a 30-second native 4K AI video model that accepts 50 reference inputs.

      **TL;DR**: At its Beijing conference, ByteDance revealed Seedance 2.5, a video generation model that can create 30-second native 4K videos using up to 50 multimodal inputs. This model was presented as a significant upgrade, skipping over four intermediate versions and moving directly from its predecessor. An enterprise beta is currently available, with a public release planned for early July. CEO Liang Rubo emphasized that advancing in AI is the company's highest priority, transitioning its model-as-a-service into a long-term foundational operation.

      Seedance 2.5 enhances its reference capacity, now allowing inputs such as images, audio, 3D white models, and style references, which improves control over style, motion, and composition compared to its predecessor's 12 inputs. The model generates native 4K output instead of upscaling, which is crucial for professional workflows, and supports 10-bit color depth for better gradients and post-production grading. ByteDance asserts that prompt adherence has improved by 20%, resulting in fewer attempts for a usable output.

      Additionally, the model processes audio in the same latent space as visual elements, allowing for synchronized onscreen actions with corresponding sound effects. A new 3D white-box preview feature enables creators to generate low-fidelity animations before finalizing high-quality renders, positioning the model as a professional production tool rather than merely a novelty.

      This announcement follows a previous need to implement watermarking and IP protections for Seedance 2.0 after legal issues with major studios like Disney and Netflix. After pausing its global rollout in mid-March, ByteDance resumed with added features to comply with copyright regulations, although no timeline for the new model's availability in the U.S. has been specified.

      Since February, the competitive landscape has changed significantly, with OpenAI shutting down its Sora tool, which experienced high operational costs and limited revenue. Google’s Veo 3.1 has stepped in, offering similar capabilities but with fewer reference inputs than ByteDance's new model.

      The AI video generation market is rapidly evolving, with Chinese models advancing production tools more quickly than their Western counterparts. Despite this, whether the new model can enter global markets without sparking renewed copyright disputes remains uncertain. ByteDance has the infrastructure and user base through CapCut but still lacks a resolution with Hollywood, with each enhancement to the model increasing the stakes of these unresolved issues.