Luma AI launched Dream Machine (powered by Ray 1) with immediate open web access, generating 5-second physically coherent text-to-video and image-to-video clips with keyframe extension months before Sora became publicly available.
Luma Dream Machine & Ray
Also known as: Luma AI · Dream Machine · Luma Ray 2 · Luma Ray 3
Plain English
Luma AI's family of video and image foundation models—spanning the original Dream Machine and the Ray 2 and Ray 3 video architectures built for rapid, physically coherent motion and HDR studio production.
Technical Overview
Luma's Ray lineage trains multimodal spacetime Diffusion Transformers directly on joint video-text-image representations. Ray 2 scaled training compute by 10x over the original Dream Machine for multi-character motion physics, while Ray 3 introduced multimodal reasoning chains, native 16-bit HDR / ACES EXR color pipelines, and multi-keyframe temporal control.
Timeline (4)
2024
Release2025
ReleaseLuma AI released Ray 2, a multimodal spacetime DiT trained with 10x the compute of Ray 1 to produce natural human locomotion, fluid physics, and complex camera trajectories.
- Release
Luma AI introduced Ray 3, pairing visual self-verification reasoning with native 16-bit HDR video generation and ACES EXR export for professional visual-effects grading pipelines.
2026
ReleaseLuma AI released Ray 3.2, expanding temporal conditioning up to 16 reference keyframes across a shot so directors can lock character staging and camera marks throughout longer generated sequences.
See also (8)
Open full graphLuma Dream Machine & Ray