Luma's Ray3.2 Lets Directors Pin 16 Keyframes to Control AI Video

Ray3.2 is Luma's sharpest answer yet to a question the AI video field has been circling for two years: what does it actually mean to direct a generated video, rather than just prompt one? The new model, designed in collaboration with creatives from the entertainment, advertising, and gaming industries, moves the control surface from a text box to something closer to a real production timeline.

From prompting to directing

Ray3.2 allows users to direct AI video frame by frame. With this update, creators are no longer just prompting , they are directing. That framing isn't just marketing. The headline capability is Multi-Keyframe, which lets you pin up to 16 reference frames inside a single clip. Think of keyframes here the same way an animator does: anchor points that tell the model exactly what the scene should look like at specific moments, with the model filling in the motion between them.

By enabling the placement of up to 16 keyframes within a single clip, Luma grants directors and production teams absolute authority over pacing and motion. Previously, you could set a start and end frame and hope the model interpolated something usable. Now you can choreograph a camera move, a character gesture, and a lighting change all within one generation pass.

What's actually new

• Multi-Keyframe (up to 16): Pin exact visual states at any point in the clip. Control camera paths, narrative beats, and visual progressions without re-generating.
• Expressive Facial Performance: Creators can choreograph exact narrative beats, camera paths, and visual progressions. Enhanced Performance Tracking and Expressive Facial Performance ensure that the nuance of a scene is accurately captured and transferred, maintaining skeletal posture and gestures, while tracking the full expressive state for up to eight faces simultaneously, frame by frame.