快手发布可灵 Kling 1.0

2024 年 6 月 6 日前后，快手把“可灵”三个字推到创作者面前时，没有先讲一套新的注意力公式。它讲运动：人物能不能跑起来，镜头能不能推拉，一秒里有没有足够多的“像拍过”的东西。发布材料里反复出现的词是大幅运动、镜头语言与较长镜头——这些词来自片场，不来自论文摘要。

Sora 的演示已经改写了公众对视频生成的胃口。问题不再是“有没有实验室能做”，而是“谁能让普通人排队用上，并且镜头别在第三秒碎掉”。中国短视频平台坐在镜头数据上十年，却长期把模型能力藏在推荐与特效里。可灵把能力翻到台前，名字直接叫生成。它开放的是文生视频与图生视频两条入口，面向创作者与企业分批试用，而不是一次性向所有人敞开。

公开材料强调较大运动幅度与更长镜头，这些句子必须带着产品边界读：具体上限、排队策略与审核规则随版本变，基准也多来自厂商设定。真正进入史册的，是坐标变了——讨论全球视频模型时，中国平台产品成为必须点名的一列。此前的对照表上，文生视频的默认名字来自旧金山与伦敦；可灵之后，北京与深圳的团队也拿到了入场券。

对广告与短视频团队，意义很具体：分镜不再只能手绘或实拍试错，可以先生成一段“像成片”的运动草稿。伪影仍在，身份仍会漂，但工作流的第一公里被改写。Runway、Luma、后来的 Seedance 与 Vidu，从此与可灵落在同一张对照表上。竞争也从“谁能生成”转向“谁能在真实产品里反复生成而不崩”。

可灵不是证明中国“突然会做视频”。它证明数据与分发原本就在这里的公司，终于选择用模型产品说话。场记板响的地方，从摄影棚扩展到了推理集群。对后来者而言，这条路的启示同样直接：视频模型的胜负手，一半在架构，另一半在十年镜头数据里。

Around 6 June 2024, when Kuaishou put “Kling” in front of creators, it did not open with a new attention formula. It talked about motion: whether people could run, whether cameras could push and pull, whether a second contained enough that felt filmed.

Sora had already rewritten public appetite for video generation. The question was no longer whether a lab could demo, but who could let ordinary people queue for something that did not shatter by second three. Chinese short-video platforms had sat on camera-language data for a decade while keeping model power inside ranking and effects. Kling flipped that power outward and named it generation.

Public materials stressed text-to-video and image-to-video, larger motion, and longer shots. Those sentences need product boundaries: caps, queues, and moderation move by version, and benchmarks are often vendor-set. What enters the chronicle is the changed coordinate system—when the world discusses video models, a Chinese platform product becomes a required column.

For ads and short-video teams the change is concrete: storyboards need not begin only as sketches or live trials; they can start as motion drafts that look nearly final. Artifacts remain; identity still drifts; but the first kilometer of the workflow is rewritten. Runway, Luma, later Seedance and Vidu now share one comparison sheet with Kling.

Kling does not prove that China “suddenly learned video.” It proves companies that already held data and distribution finally chose to speak in model products. The clapperboard now sounds in inference clusters as well as studios.