Luma Dream Machine v1.6

Informazioni su Luma AI

Fondato 2021 · Palo Alto, USA

Luma AI was founded in 2021 by Amit Jain (CEO, ex-Apple Vision/Camera) and Alex Yu (ex-UC Berkeley, NeRF research). The company originally focused on 3D capture and Neural Radiance Fields, launching the Luma iOS app for AR/3D scanning in 2022. In June 2024 Luma released Dream Machine, a text- and image-to-video generative model that became one of the first widely viral consumer video AIs alongside Runway Gen-3 and Kling. Dream Machine v1.6 (October 2024) introduced extended duration, camera control concepts and improved motion consistency. Investors include a16z, Amplify Partners, NVIDIA and Matrix Partners, with Series C funding announced in 2025 valuing the company above $4B. Luma also released the Ray2 model family for higher-fidelity video.

Architettura

Latent video diffusion model with transformer denoiser (DiT-style)

Dream Machine is a closed-source latent video diffusion model that compresses video into a learned spatio-temporal latent space and uses a transformer-based denoiser conditioned on text embeddings (from a multilingual text encoder) and optional image conditioning. Version 1.6 supports text-to-video, image-to-video (first frame), end-frame conditioning and 'extend' operations that continue an existing clip while preserving identity and motion. Luma have publicly described Dream Machine as targeting 'consistent characters, accurate physics and complex camera motions'. The training corpus combines licensed stock video, web-scale curated video and synthetic data; captioning is produced by an in-house vision-language model. v1.6 generates 5-second clips at 720p (with upscaling to 1080p in post) at 24 fps natively.

Parametri

Undisclosed (estimated multi-billion)

Capacità

• Text-to-video and image-to-video generation
• Start-frame and end-frame conditioning for guided shots
• Extend functionality to continue an existing clip up to ~20s
• 5-second base clips at 720p, upscalable to 1080p
• Strong cinematic camera moves (dolly, orbit, push-in)
• Consistent character identity across extensions
• Available via web app, API and integrations (Discord, ComfyUI)
• Lower cost per second than Gen-3 Alpha at launch
• Best for: marketing teasers, social-media reels, mood films, music videos.

Addestramento e licenza

Closed corpus combining licensed stock footage, curated web video and synthetic data; bilingual captions from an in-house vision-language captioner. Exact dataset size undisclosed.

Licenza: Proprietary commercial licence via Luma API and luma.com terms of use; generated video may be used commercially on paid plans.

Test di sicurezza: NSFW, CSAM and likeness filters at prompt and output stage; public figures filter; no formal third-party red-team disclosed.