Luma AI Guide 2026

Luma AI is the product family from Luma Labs, a San Francisco AI research company founded in 2021. Luma ships four main things in May 2026. Dream Machine is the text-to-video web app at dream-machine.lumalabs.ai. Ray 2 is the underlying video model that powers Dream Machine. Photon is the text-to-image model used inside Dream Machine. Genie is the text-to-3D mesh model at lumalabs.ai/genie. There is also a developer API at lumalabs.ai/api exposing Ray 2, Photon, and Genie endpoints. Luma raised a $43M Series B led by Andreessen Horowitz in 2024 and has been shipping monthly since.

Dream Machine has four tiers. Free at 30 daily credits with watermark and public feed. Lite at $9.99 per month (annual) for 1,600 monthly credits with Ray 2 Flash. Plus at $29.99 per month (annual) for 4,200 monthly credits with full Ray 2 and 1080p upscale. Unlimited at $94.99 per month (annual) for unlimited Relaxed-speed generations plus a fixed monthly credit pool for Fast-speed jobs. Monthly billing is roughly 30 percent more expensive. Prices verified from lumalabs.ai/dream-machine/pricing on May 18, 2026.

Ray 2 is the full quality model that produces the cleanest motion and tightest prompt adherence Luma has shipped. Ray 2 Flash is a smaller, cheaper variant introduced October 8, 2025 that costs roughly one third of the credits per clip in exchange for slightly noisier motion and weaker handling of complex compositions. Flash is the default on the Lite tier, full Ray 2 unlocks on Plus and above. For social-cut work and exploration, Flash is fine. For client deliverables, jump to full Ray 2.

Dream Machine with Ray 2 produces 10 second clips at 1080p with strong natural-motion physics. Sora 2 reaches 20 second clips but is locked to ChatGPT Plus or Pro subscribers and is geo-restricted. Google Veo 3 caps at 8 seconds but is the only major model that generates synchronized native audio in a single pass. The deciding factor between Luma and the others is usually motion realism. Ray 2 handles physics (falling objects, water, fabric) more believably than any other consumer video model as of May 2026. If you want native sound, go to Veo. If you want long single clips, go to Sora.

Extend lets you take a Dream Machine clip and add another 5 seconds (or another 10 seconds on Ray 2) to either end. The model uses the final frame of the prior clip as the implicit starting frame and continues the motion from there. You can chain extensions to build a 30 second sequence on Ray 2, although coherence loosens after the second extension. Extend uses standard credits, the same as a new generation of equivalent length, so a 30 second chain costs roughly three times a 10 second clip.

Genie is Luma's text-to-3D mesh generator at lumalabs.ai/genie. You type a prompt and Genie returns a downloadable GLB, OBJ, or USDZ mesh in roughly 10 seconds. As of Genie 1.5 the output quality is good enough for game prototype assets, AR placeholder objects, and Blender starting meshes, but not production-grade game assets without manual cleanup. Topology is messy on organic shapes, and texture maps are baked rather than separated. For studio-grade 3D, you still need a human modeller. For previs and concept iteration, Genie is genuinely useful.

Yes. The Luma API launched September 23, 2024 at lumalabs.ai/api and exposes endpoints for Dream Machine video, Photon image, and Genie 3D. Pricing is metered. Ray 2 Flash is roughly $0.10 per 5 second clip, full Ray 2 is roughly $0.40 per 5 second clip, Photon Flash is roughly $0.002 per image, and Genie is roughly $0.05 per mesh. There is a free $5 trial credit on signup. The Luma API is also resold through fal.ai, Replicate, and PiAPI if you prefer a single aggregator endpoint.

Commercial rights are granted on the Lite, Plus, and Unlimited tiers per the Luma Terms of Service. The Free tier explicitly forbids commercial use and posts your generations to a public feed at dream-machine.lumalabs.ai/feed. Luma grants you the right to use, modify, and distribute outputs subject to the standard content policies (no minors, no real political figures in fabricated scenes, no identifiable likeness without consent). The watermark is removed on all paid tiers.

Three common causes. First, your prompt has multiple subjects in motion at the same time. Two characters interacting still trips Ray 2 in roughly one in three generations. Second, you used Ray 2 Flash on a complex composition. Upgrade to full Ray 2 for anything with detailed faces or hands. Third, your starting image (on image-to-video) has fine detail that the model cannot preserve at the upscale step. Reframe to a closer shot or simplify the background. Luma in our testing was strongest on single-subject scenes with one motion verb.

Yes, and it is the workflow most experienced Dream Machine users default to. You upload a still image, write a one-sentence motion prompt, and let Ray 2 animate it. This approach gives you tighter compositional control than pure text-to-video because the model is anchored to your starting frame. The trade-off is that very stylized starting images (cel-shaded anime, heavy oil paint) sometimes lose stylistic consistency mid-clip. Photoreal starting images stay coherent more reliably.

Dream Machine officially accepts English prompts. In practice it handles Spanish, Portuguese, French, German, Japanese, and Korean acceptably for simple prompts, with noticeably weaker adherence on prompts longer than 50 words in non-English languages. For multilingual creators, the recommended workflow is to draft the idea in your native language, then translate the final 30-to-50 word prompt to English before submitting. Chinese prompts work but lag the model's behaviour for Chinese cultural references compared with Kling, which is built by Kuaishou.

Yes for storyboards, animatics, social cuts, music videos, mood reels, and pitch decks as long as you are on a paid tier so you have commercial rights. We would not use Ray 2 outputs as final hero footage on a TV spot or theatrical release without VFX cleanup. Text rendering, hands, and complex multi-subject interactions still show artifacts. The motion realism is the standout reason to pick Luma over Kling or Runway. The lack of native audio is the reason to pair it with ElevenLabs or to switch to Veo 3 when sound matters.