TL;DR
ByteDance announced Seedance 2.5 at its Beijing conference, which generates 30-second native 4K video from up to 50 multimodal reference inputs.
ByteDance unveiled Seedance 2.5 on Tuesday at its Volcano Engine FORCE conference in Beijinga video generation model that produces 30-second clips at native 4K resolution from a single message. The company skipped four intervening versions entirely, jumping directly from its predecessor to signal what it described as a generational leap.
An enterprise beta version is now available, with a public launch planned for early July. CEO Liang Rubo said at the conference that rising to the top of AI is the company’s top priority, with its as-a-service model business evolving into a core operation backed by long-term investments.
The main upgrade is the referencing capability: the model accepts up to 50 multimodal inputs, including images, audio clips, 3D blank models, and style references, up from 12 for its predecessor. Those inputs give Seedance 2.5 much more granular control over style, movement, and composition than a text message alone.
The model outputs in 4K natively rather than upscaling from a lower resolution, a distinction that’s important for professional production pipelines. Supports 10-bit color depth for smoother gradients and more room for color grading in post-production. ByteDance also claims to have 20 percent better fast adhesion, meaning fewer generations before you get a usable result.
Audio is now co-processed within the same latent space as visual signals, producing native synchronization between on-screen actions and their corresponding sound effects. A new 3D white box preview feature allows creators to generate low-fidelity animations before committing to a full-quality render. Together, the features position the model as a production tool rather than a novelty generator.
The announcement comes three months after ByteDance was forced to add watermarks and IP firewalls to Seedance 2.0 following cease and desist letters from Disney, Warner Bros Discovery, Paramount and Netflix. A viral deepfake of Tom Cruise fighting with Brad Pitt on a rooftop prompted a formal complaint from the Motion Picture Association and a reprimand from SAG-AFTRA.
ByteDance paused the global rollout in mid-March and did not resume it through CapCut until late March, with face blocking filters, C2PA watermarking, and copyrighted character detection. No timeline has been offered for the new model to be available in the United States.
The competitive context has changed drastically since February. OpenAI shut down Sora in March after the video tool peaked at around one million users and reportedly cost around $1 million a day to run, generating just over $2 million in total revenue.
Google’s Veo 3.1 has filled much of the gap, offering native 4K output, audio generation, and up to three reference images for styling control. But the new ByteDance model substantially surpasses Veo’s baseline input capacity, accepting 50 inputs versus Veo’s three, a gap that’s important for professional workflows.
The AI video generation market has rapidly fragmented, with Chinese models advancing faster in production tools than Western competitors. Third party platforms such as Reallusion’s AI Studio has already created professional channels around the predecessor model.and Runway’s fourth generation tool has left the top 10 of Artificial Analysis.
The central question remains whether the new model can reach global markets without reigniting the copyright battles that paralyzed its predecessor. ByteDance has the model, distribution across CapCut’s 400 million monthly active users, and vertical integration from generation to publishing to sharing. What it doesn’t have yet is a deal with Hollywood, and every feature that makes the model more capable also increases the risks of that unresolved conflict.
