ByteDance's Seedance 2.5 video AI handles 50 references, 30-sec 4K clips

ByteDance has quietly shown off Seedance 2.5, the latest iteration of its AI video generation model, at a Beijing conference last month. According to reports from The Information and Testing Catalog, the model could launch as early as July 9. A landing page is live on the Dreamina website, and CapCut tweeted about an impending release on July 4. The model's standout feature: users can now provide up to 50 reference pieces — images, videos, or audio files — up from 12 in Seedance 2.0, giving far greater creative control. It can output 30-second clips at 4K resolution from a single text prompt, putting it toe-to-toe with rivals like OpenAI's now-defunct Sora and Google's Veo 3.

However, the rollout may face headwinds. Seedance 2.0's US launch was delayed earlier this year after Hollywood pressure over alleged use of copyrighted works for training. If Seedance 2.5 is significantly more capable, it could renew scrutiny. The model also raises deepfake concerns; the rapid improvement in AI video realism makes watermarking critical. ByteDance, which owns a majority stake in TikTok, has not commented on the launch date or copyright issues. For now, the industry watches as another major player in generative video prepares to enter the fray.