Luma 发布 Ray2 视频模型

一个人在暴雨里跑过街口，跳上公交车。静止帧可以很漂亮，真正暴露视频模型的却是中间那几秒：腿会不会突然错位，雨伞是否穿过车门，镜头转动后还是不是同一个人。Dream Machine 上线初代模型时，用户已经能够很快得到一段视频；到了 2025 年 1 月 15 日，Luma 用 Ray2 回答的是下一层问题——这段运动能不能被人相信。

Ray2 首先进入付费订阅用户的产品界面，提供文生视频与图生视频。可选片段大约为 5 秒或 9 秒，并有 540p、720p 等分辨率设置。这些规格并不宏大，却直接规定了创作者怎样工作：九秒仍不足以完成一个场景，但足够做一条镜头草稿；720p 未必是最终交付，却能让团队判断动作、构图和节奏是否值得继续。

Luma 称 Ray2 使用了约前代十倍的训练算力，并强调快速、连贯的运动，更多细节，以及更合理的事件顺序。十倍算力只能说明投入规模，不能证明每个提示都提升十倍。用户真正能量出的，是同一镜头要重生几次：如果角色少变一次脸、物体少穿模一次，模型就替剪辑师省下了一轮挑选和补救。

这也是当时视频产品竞争变化最快的地方。Sora 抬高了演示片的上限，Runway、可灵等产品则让创作者习惯在可购买的工具里反复生成。Luma 必须让“更自然的运动”出现在订阅账户中，而不只是发布视频里。后来进入云平台分发，也说明 Ray2 不只想做 Dream Machine 的一个按钮，还要成为其他应用可以调用的生成能力。

Ray2 没有消灭伪影，更没有学会现实世界的完整物理。它把失败的位置往后推了一点：从第一眼就不对，推到需要多看几秒才发现。对视频创作，这一点距离很实在。镜头仍然短，成片仍然需要剪辑，但生成内容开始有机会从灵感展示进入预演、提案和素材草稿。

A person runs through a rainstorm, crosses an intersection, and jumps onto a bus. Any individual frame can look convincing. The seconds between them reveal the model: a leg may change direction, an umbrella may pass through the door, or the person after the camera move may no longer be the person who began the shot. Dream Machine’s first generation had shown that users could obtain a video quickly. With Ray2 on January 15, 2025, Luma addressed the next question—could viewers believe the motion connecting those frames?

Ray2 first appeared inside the paid Dream Machine experience for text-to-video and image-to-video work. Creators could generate clips of roughly five or nine seconds with settings including 540p and 720p. Those specifications were modest, but they shaped the job precisely. Nine seconds was not a complete scene, yet it was enough for a shot draft. A 720p result might not be final delivery, but it could tell a team whether the action, framing, and timing deserved another pass.

Luma said Ray2 had been trained with roughly ten times the compute of its predecessor and emphasized fast coherent motion, greater detail, and more logical sequences of events. The compute multiplier described investment, not a tenfold guarantee for every prompt. Creators had a more practical measurement: how many generations did the same shot require? If a character changed faces one fewer time or an object avoided one impossible collision, the model removed a round of selection and repair from the edit.

That was where competition among video products was moving fastest. Sora had raised expectations through demonstrations, while products from Runway, Kling, and others taught creators to regenerate inside tools they could actually buy. Luma needed “more natural motion” to exist in a subscriber’s account, not only in a launch reel. Ray2’s later distribution through cloud platforms also suggested an ambition beyond a button in Dream Machine: the model could become a capability embedded in other applications.

Ray2 did not eliminate artifacts or learn a complete simulation of the physical world. It pushed the visible failure slightly later—from something wrong at first glance to something a viewer might notice only after watching for several seconds. In video work, that distance was useful. Clips remained short and finished pieces still depended on editing, but generated material had a better chance of moving from inspiration demo to previz, pitch, or working shot draft.