Ray3.2, Luma's video model that combines four workflows and keyframe control

Luma launches Ray3.2, a cinematic video model with four workflows and 16-keyframe control, available via Dream Machine and API for professional production.

Luma makes Ray3.2 its most versatile video model, presented as a clearly cinema production-oriented evolution compared to Ray 3 and Ray 2, with more control, continuity, and cinematic direction. The tool combines four workflows: Modify Video, image2video, text2video, and Reframe. The most notable change concerns control: where previous versions started from a single initial image (sometimes a final one) and generated everything in between on their own, Ray3.2 allows up to 16 keyframes per clip, enabling precise decisions on where light shifts, where a movement settles, and where a narrative beat plays out.

Modify Video V2 reworks an existing sequence, preserving performance, movement, and timing while restyling the image: users can change the setting, lighting, or wardrobe, with expressive facial tracking for up to eight faces and performance tracking that maintains skeletal pose and gestures throughout the duration. Length increases up to 20 seconds at 1080p, depending on the source frame rate (20s at 24 fps, 15s at 30 fps, 7s at 60 fps). Reframe, for its part, adapts a video to all formats (from 9:16 to 21:9) by filling the frame with context consistent with the scene.

On the production side, the model generates in 360p, 540p, 720p, and 1080p, outputting in .mp4 but also in native 16-bit HDR and EXR (ACES2065-1, ready for color grading and compositing). text2video and image2video do not produce native audio, which is added in a separate step, while Modify Video and Reframe retain the original track. Everything is accessible from Dream Machine, and a Ray3.2 API opens the same range of controls (multi-keyframe, performance tracking, EXR export) to studios, agencies, and developers.