Luma Ray 3.2 released: returning "directorial control" of video generation to creators

The most commonly criticized point of the video generation model is that it is "uncontrollable" - the picture is beautiful, but the camera does not obey the command. Luma’s Ray 3.2, released on June 9th, directly targets this pain point: it is officially defined as “putting every frame under your control, maintaining continuity across cuts, and finishing with cinematic quality.” Translated into vernacular, it allows users to operate video generation like a director.

From "generating images" to "directing videos"

The core capabilities of Ray 3.2 revolve around three points: Accurately control each frame, Maintain subject and style consistency across shots, and Output movie-level films. It is no longer a tool for "writing prompt words and other results", but a creative process of "checking frame by frame and directing across lenses". With the Ray series API, creators can embed this set of control capabilities into their own workflows.

It is worth emphasizing the technical background of Luma: the official has repeatedly emphasized that "it is not a wrapper, we build the bottom layer of the model." Before Ray 3.2, Luma had underlying accumulation such as Terminal Velocity Matching (efficient inference scaling, engineering report in November 2025), Inductive Moment Matching (breaking through the ceiling of pre-training algorithms, March 2025). The Ray 3 evaluation report released in October 2025 also claimed that SOTA professional video generation-Ray 3.2 is the successor of this self-research line.

The core position in the three-tier architecture

Putting Ray 3.2 into the Luma 2026 product landscape, the position is very clear: