字节跳动推出AI视频生成工具Seedance 2.5，在视频生成能力、素材支持和编辑精度上实现多项升级[1]。相比2.0版本，新版本支持原生30秒长视频一镜到底生成（非多段拼接），参考素材支持量从12个扩充至50个，包括30张图片、10段视频和10段音频[1]。此外，Seedance 2.5新增白模参考功能，可精准控制空间结构、人物姿态和镜头机位，同时支持局部场景编辑和时间戳精确控制[1]。
该工具已进入B端生产领域[1]。穹彻智能、微分智飞、徐工集团等企业已将Seedance应用于具身智能、飞行机器人、工业培训和自动驾驶等场景[1]。火山引擎总裁谭待表示，Seedance是字节AGI战略的核心基础设施，视频生成是构建世界模型的基础[1]。他同时指出，Seedance现有定价偏低，覆盖算力成本都困难[1]。
ByteDance has unveiled Seedance 2.5, an upgraded version of its artificial intelligence video generation tool that introduces significant technical enhancements and expands its applications beyond consumer use cases [1]. The new version enables native generation of 30-second continuous videos, a substantial improvement over previous capabilities [1]. The system now supports up to 50 multimodal reference materials—comprising 30 images, 10 video clips, and 10 audio segments—compared to 12 in the 2.0 release [1].
Additional upgrades include white-model referencing for precise spatial structure and character pose control, as well as localized scene editing with timestamp-level precision [1]. According to Tan Dai, president of VolcanoEngine (ByteDance's cloud division), the pricing of Seedance remains challenging given production costs, as current rates struggle to cover computational expenses [1]. Tan characterized video generation as a foundational capability for constructing world models and identified Seedance as a core strategic instrument in ByteDance's broader AGI initiative [1].
Beyond consumer applications, Seedance 2.5 has already been deployed across multiple enterprise sectors [1]. Adoption spans embodied AI systems at Qiongche Intelligence, autonomous flying robots at Weifen Zhifei, and industrial training programs at Xuzhou Construction Machinery Group [1].