Seedance 2.0 发布与版权风暴

Seedance 2.0 的官方发布页展示了一套比 1.0 更像制作台的输入方式：文字、图片、音频和视频可以混合提供，模型再根据构图、动作、镜头、声音或特效生成新片段。用户最多可以同时提交九张图片、三段视频和三段音频；输出支持十五秒多镜头音画，并加入双声道声音。2026 年 2 月 12 日，字节发布的不再只是“画面会动”的模型，而是一套试图接管更多制作步骤的系统。

这组规格也解释了为什么 2.0 的样片更容易越过公众的心理边界。角色可以带着熟悉的脸和服装进入新的镜头，参考视频可以提供动作与摄影语言，生成的对白、环境声和音乐又让片段更接近已经剪好的内容。字节强调复杂多人运动、物理表现和可控性，同时在自己的发布材料中承认结果仍有缺陷，音频失真、长期一致性与物理规律都没有彻底解决。

用户传播把技术发布迅速改写成权利事件。社交网络出现大量明星、电影角色和可辨认场景的写实短片；工作室与行业组织在数日内发函或公开表态。争议并不只指向某一张生成图片，而是同时追问训练材料来自哪里、产品是否允许复刻受保护角色，以及平台在输出已经高度相似时应当怎样阻止传播。

字节回应称尊重知识产权并将加强防护。到 3 月，全球发布计划被报道暂停。这里必须保留措辞边界：发函不是判决，暂停全球上线也不等于所有已有入口同时关闭。但对产品团队而言，效果已经发生。过滤策略、地区开放和参考素材规则从上线后的治理工作，提前成了决定模型能否上线的条件。

Seedance 1.0 让字节拥有一个可被比较的视频基座；2.0 则让这个名字承担了平台责任。能力越接近成片，创作者越容易获得有用素材，也越容易在几分钟内制造一段权利关系复杂、可以大规模传播的内容。模型的“可用率”不再只由伪影多少决定，还取决于生成结果是否能被合法地保存、发布和商业使用。

这场冲突没有在发布周得出最终答案，却改变了后来每一次视频模型发布的检查顺序。人们会先看画面，再问参考素材能否上传、角色是否被拦截、哪些地区有入口。Seedance 2.0 把版权从产品条款的末尾，推到了生成按钮之前。

The official Seedance 2.0 launch described an input surface closer to a production desk than the one used by its predecessor. Text, images, audio, and video could be supplied together, with the model referring to composition, action, camera movement, sound, or effects when it created a new clip. A request could include as many as nine images, three videos, and three audio clips. Output extended to fifteen-second multi-shot audiovisual sequences with two-channel sound. On February 12, 2026, ByteDance was no longer releasing only a model that made pictures move. It was proposing a system that could absorb more of the production process.

Those specifications also help explain why Seedance 2.0 outputs crossed a public threshold so quickly. A recognizable face and costume could enter a new scene, a reference video could supply movement and camera language, and generated dialogue, ambient sound, and music made the result resemble edited media. ByteDance emphasized complex multi-person motion, physical accuracy, and control. Its own launch material also acknowledged remaining flaws: audio distortion, long-term consistency, and adherence to physical rules had not been eliminated.

User distribution rapidly rewrote a technical release as a rights event. Social feeds filled with photoreal clips of celebrities, film characters, and recognizable scenes. Studios and industry organizations sent letters or issued public objections within days. The dispute was not limited to one generated image. It asked where training material came from, whether the product permitted replication of protected characters, and what a platform should do when an output was similar enough to circulate as familiar media.

ByteDance said it respected intellectual property and would strengthen safeguards. By March, plans for a global release were reported suspended. The wording matters. A legal letter was not a judgment, and a paused global rollout did not mean that every existing point of access disappeared at once. The product effect was nevertheless immediate. Filters, regional availability, and rules for reference material moved from post-launch governance into preconditions for whether the model could ship.

Seedance 1.0 had given ByteDance a video foundation model that could be compared with its peers. Version 2.0 made the same name carry platform responsibility. As capability moved closer to finished media, creators could obtain more useful material—and could produce, within minutes, content with complicated rights that was ready for mass distribution. “Usability” could no longer be measured only by artifact rates. It also depended on whether an output could legally be stored, published, and sold.

The conflict did not produce a final answer during launch week. It changed the order of questions asked at later video-model releases. Viewers would inspect the image, then ask which references could be uploaded, which characters were blocked, and which regions had access. Seedance 2.0 moved copyright from the end of the terms of service to a position in front of the generation button.