The site presents itself in one sentence: Luma is your creative partner, with the idea yours and the realisation handed over. The mission statement goes further, describing a push toward a unified general intelligence that can generate content, understand the world and act inside physical space. The naming history is worth a note. A migration notice on the API pages still refers to a legacy video interface under the old Dream Machine name, while the current line is Ray and UNI-1, threaded together by Luma Agents.

Ray 3.2 puts every frame under direction

The product page leads with its claim: direct any frame, finish every cut. The official wording says the model puts each frame in the user's hands, keeps continuity between cuts and delivers at cinema quality. It covers text to video, image to video and video to video, plus a reframing mode, and 1080p output appears across all of them.

Up to sixteen keyframes in one clip

The most practical control is the multi-keyframe feature. A single clip can hold as many as 16 keyframes, and the user decides what changes, what holds still and how the story closes. The same lever is exposed through the interface, so it can be written into an automated pipeline. Around control the page lists motion transfer, camera motion transfer, character swap, visual effects, environment replacement, relighting and product replacement, all of them in the family of keeping the original performance while rebuilding the image.

Rewriting footage you already shot

The second generation of the modify tool does something the industry needs badly. It rewrites material you already own, swapping walls, environments and clothing while keeping the lighting consistent and the performance intact, up to 20 seconds at a time and entirely in 1080p. That answers the most common rework problem in film and advertising, where a reshoot costs too much but the image has to change. The page groups it with reframing, improved motion transfer and expressive facial performance under one idea: create once, deliver everywhere.

Output that a finishing house will accept

This is where the model separates itself from most video generators. It supports native HDR generation, with 16-bit colour produced by the model rather than stretched in post, and it exports 16-bit EXR frames in the ACES2065-1 colour space with the AP0 primaries, which the page says means it drops straight into compositing. The API documentation is plainer still: native HDR and EXR export exist so that generated material can enter a grading or compositing pipeline alongside live action. Pricing follows the same logic, with standard dynamic range at the base rate, HDR at double and HDR with EXR at triple.

The image model, and the agents that string it together

On the still-image side sits UNI-1, pitched as brand understanding pushed down into the model itself: it reads a brand as deeply as it generates, carrying style references, character references and visual benchmarks into every output. A follow-up release is sold on half the price, half the latency and getting it right the first time. Agents are the connective tissue, running research, generation and refinement across video, image, audio and text while holding the brand tone steady throughout.

Three personal tiers, starting at 10000 credits

Personal subscriptions come in three tiers, priced in US dollars, with up to 20% saved by paying yearly. Plus costs 30 dollars a month, or 300 a year, and carries 10000 credits a month. Pro costs 90 dollars a month, 900 a year, with 40000 credits and four times the agent usage of Plus. Ultra costs 300 dollars a month, 3000 a year, with 150000 credits and fifteen times the agent usage. Even Plus includes access to Luma models and to third-party image and video models, guest editing rights and a commercial licence. Team and Enterprise tiers exist for organisations, quoted through sales, adding member management, pooled credits, usage analytics and enterprise commitments.

What credits actually buy

The pricing page publishes a line-by-line table. A five second clip from text or from an image costs 20 credits at draft resolution of 360p, 50 at 540p, 100 at 720p and 400 at 1080p, with ten seconds scaling proportionally. Video to video at five seconds and 1080p costs 720 credits. Reframing is billed by the second at 120 credits per second in 1080p. Run the arithmetic and the 10000 credits on Plus come to roughly 25 finished five second clips at 1080p.

The interface is billed per call

Anyone putting generation inside their own product meets usage pricing. The published reference for the Ray model puts a five second text or image clip at 15 cents in 540p, 30 cents in 720p and 1 dollar 20 cents in 1080p, with ten seconds at 45 cents, 90 cents and 3 dollars 60 cents, and video to video at five seconds costing 72 cents, 1 dollar 44 cents and 2 dollars 16 cents. Two plans exist: one is pay as you go with no minimum, aimed at prototypes and early production, the other offers dedicated capacity with throughput and latency guarantees, a service agreement, content moderation and a promise that customer material is not used for training.

Three boundaries

Credits go quickly: 400 for five seconds at 1080p means Plus stretches to about 25 clips a month, and video to video plus HDR or EXR output multiply that again, so test parameters at draft resolution first. What the site offers is a free trial rather than a permanent free tier, personal plans begin at 30 dollars a month, and team or enterprise pricing has to be quoted by sales. Interface prices are approximations converted from billing units, so the current published rate governs settlement, and commercial use should be checked against the licence and the content rules before anything ships.

The point of Luma is not another doorway for turning text into video. It is control taken to an engineering standard: 16 keyframes in a clip, footage rewritten for up to 20 seconds, native HDR and 16-bit EXR on the way out, all so that generated material can enter a post-production pipeline that already exists.