Luma Dream Machine.
The 3D-native video model — born from Neural Radiance Field research, now generating film.
What it is.
Palo Alto-based research lab founded by the founding team. Started with NeRF and 3D reconstruction — the camera-motion intuition shows in every Dream Machine output. Ray 3 shipped in 2025 with hi-fi 1080p generation, camera control, and the strongest spatial coherence in the video-gen market. Backed by Andreessen, Matrix, and Nvidia.
Where it fits.
Where the brief demands camera grammar. Film and TV pre-visualisation. Architectural walkthroughs. Spatial computing content for Vision Pro and Quest. Operators who think in dolly shots, pans, and pulls — not just prompts. The model trained on the largest 3D dataset in the category.
- Strongest camera control and spatial coherence
- 3D and Gaussian Splat output via Genie
- Photorealistic motion at Pro tier and above
- Premier tier pricing steep for casual users
- Output volume lower per credit than Pika and Sora
Frequently asked.
What is Luma's 3D heritage?
Luma began with NeRF reconstruction — capturing real-world scenes from phone video. The 3D priors show in Dream Machine's camera coherence.
Can Luma generate 3D assets?
Yes via Genie. Outputs include glTF and USDZ formats — usable in Unity, Unreal, and visionOS pipelines.
How does Ray 3 compare to Sora?
Ray 3 leads on camera motion and 3D coherence. Sora leads on character motion and physics. Most pre-vis teams run both.
