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

      TL;DR: ByteDance introduced Seedance 2.5 at its conference in Beijing, capable of generating 30-second native 4K videos from up to 50 multimodal reference inputs. The model represents a significant upgrade from its predecessor, skipping four intermediate versions to emphasize a major advancement.

      An enterprise beta is currently active, with a public release expected in early July. CEO Liang Rubo emphasized that reaching the pinnacle of AI technology is the company’s primary focus, and its model-as-a-service business is developing into a core operation underpinned by long-term investment.

      The key enhancement involves an increase in reference capacity: the model can now accept up to 50 multimodal inputs—including images, audio clips, 3D white models, and style guides—compared to 12 with the previous version. This expanded input capability allows Seedance 2.5 to achieve more precise control over style, motion, and composition than just using text prompts.

      The model generates videos in native 4K rather than upscaling from a lower resolution, which is crucial for professional production processes. It also supports 10-bit color depth for smoother gradients and enhanced post-production color grading. ByteDance claims a 20% improvement in prompt adherence, resulting in fewer iterations needed to produce a satisfactory outcome.

      Audio is co-processed alongside visual signals, ensuring that sound effects are synchronized with on-screen actions. A new 3D white-box preview feature enables creators to generate low-fidelity animations before finalizing high-quality renders. These features position the model as a production tool rather than merely a novelty.

      This announcement follows a three-month period during which ByteDance had to implement watermarking and intellectual property protections for Seedance 2.0 in response to cease-and-desist letters from major studios after a viral deepfake incident. Global rollout was temporarily halted in mid-March and resumed at the end of the month with new safeguards. There is no specified timeline for the new model’s availability in the United States.

      The competitive landscape has changed notably since February. OpenAI discontinued its Sora video tool in March after it reached about one million users but was reported to be expensive to operate. Google’s Veo 3.1 has emerged as a significant competitor, offering native 4K output, audio generation, and three reference images for style control, but ByteDance’s model greatly surpasses Veo in reference input capacity.

      The AI video generation sector has rapidly fragmented, with Chinese models advancing production tools faster than Western alternatives. Third-party platforms have already developed professional pipelines around the predecessor model, while Runway’s fourth-generation tool has fallen out of the top rankings in Artificial Analysis.

      The crucial issue remains whether the new model can enter global markets without reigniting the copyright disputes that hindered its predecessor. ByteDance possesses the model, a distribution network through CapCut’s 400 million monthly active users, and end-to-end integration from generation to editing and sharing. However, it has yet to resolve its issues with Hollywood, and each new capability potentially complicates this unresolved conflict.