Luma launches Scenes for storyboard-first AI video

Luma has launched Luma Scenes, a workflow that has users build and approve a visual storyboard before any video is rendered. The company describes the problem as a "tax on AI video": teams often need 10 or more generations to get one usable spot, and they pay full render cost before they know whether a clip works.

Scenes takes a brief, brand reference images, or both, and uses Luma's Uni-1 model to produce a sequenced storyboard of editable keyframes. Luma says it generates all scenes together in one context, so the look, environment, and spatial logic stay consistent across the sequence. Users can regenerate a keyframe, adjust it with comments, or replace it without changing frames they have already approved. They can also drag keyframes to change timing.

Nothing renders until the user approves. Final output uses either Luma's Ray 3.2 or Seedance 2, in 9:16 vertical or 16:9 horizontal. Luma notes that the two models behave differently: Ray 3.2 tends to return several short clips, while Seedance 2 often returns one longer clip. Exports are available as a full clip, as separate clips, or as Ray 3.2 panels.

Luma is pitching the tool at commercial volume work. An approved sequence can serve as a base for variants, such as re-running the same spot for a new market with one variable changed. The company says that fixing one scene costs one scene, not the whole spot.

Scenes is live now. Luma did not give pricing, per-render costs, or benchmark data in the announcement.