Luma AI Prompts

Luma AI is a San Francisco company founded in 2021 by Amit Jain and Alex Yu. The team originally focused on phone-based 3D capture (their NeRF apps went viral), and the product line in 2026 spans Dream Machine for video, Ray 2 as the underlying video model, Photon for text-to-image, Genie 1.5 for text-to-3D, and a Luma API for developers. That breadth is the point: Luma is one of the few labs shipping competitive video, image, and 3D output from a single brand.

Dream Machine is the web app at dream-machine.lumalabs.ai. Ray 2 is the underlying video model that powers Dream Machine, launched March 31, 2025. Most users only interact with Dream Machine and never think about the model name. The reason the distinction matters is the Luma API, where you address the model directly. If you see Ray 2 or Ray 2 Flash referenced anywhere, that is the engine behind the Dream Machine clip you would otherwise generate in the web app.

Ray 2 Flash is Luma's cheaper, faster variant. It trades a measurable amount of detail for speed and cost. Pick Flash when you are iterating: trying out a composition, exploring camera moves, refining a prompt. Switch to full Ray 2 when the composition is locked and you want the final clip. Treating Flash as a draft tier and Ray 2 as a final tier keeps monthly credit consumption sane while you find the shot.

Generate a clean still first, then animate it. Either generate the still in Photon inside Dream Machine, or upload your own. From there, switch to image-to-video, write a short description of the action and the camera, and run Ray 2 Flash for the first pass. Once the composition holds, regenerate at Ray 2 quality. This loop keeps text-to-video frustration low because you have already locked the look before the model has to interpret a long prompt.

Dream Machine generates short clips, typically five seconds, with the option to extend by feeding the final frame back into the model as a new starting point. Extensions can be chained, although each extension softens the original composition slightly because the model is reseeding from a generated frame rather than the source idea. For anything longer than ten seconds, the better path is to plan two or three separate shots, generate each, and edit them together in a video editor.

Photon is Luma's in-house text-to-image model, launched November 18, 2024, with a faster Photon Flash variant. Inside Dream Machine you can use Photon to generate the still you intend to animate, which keeps your image and video in the same visual family. Photon is also exposed on the Luma API for image generation at scale. It is not aimed at competing with the strongest text-renderers; it is aimed at giving Luma users a native still pipeline.

Genie returns a low-poly 3D mesh from a text prompt: usually a couple of thousand polygons, a single connected mesh, no rigging, no textures beyond a simple base material. Treat the output as a blockout, then retopo and texture in Blender or Substance. Genie is not a finished asset pipeline, but it remains the quickest path from an idea to a manipulable mesh you can refine. For game-ready props and crowd assets, it pays back faster than starting from a primitive.

Name one action and one camera move per clip. The pattern that works: subject (4 to 8 words), action (1 verb plus 1 detail), camera (move plus angle plus lens), light (1 phrase), length. A clip like a dog walks toward the camera through tall grass, slow zoom in, soft daylight, 5 seconds reads cleanly. Stacking two actions, or describing the entire scene before the camera move, makes Ray 2 hedge and the motion goes soft. Less is more.

Yes, image-to-video accepts uploads and that mode is the most reliable way to use Dream Machine for finished work. A locked still gives Ray 2 a clear anchor for the look, and the model can focus on the motion and the camera rather than reinterpreting your subject. Photoreal product shots, on-brand illustrations, and character references all benefit from image-to-video over text-to-video. If a reference image exists, upload it instead of writing a paragraph.

Paid Dream Machine plans grant commercial use of the outputs, with the exact terms in Luma's policies at lumalabs.ai. Free tier outputs typically carry restrictions (status as of May 2026). Practically that means client work, paid social, and product launches should run through a paid plan. Re-check the current terms before a commercial shoot, since Luma updates its commercial policy as the API and enterprise tiers grow. The same caveat applies to underlying providers whose models appear inside Dream Machine.

Veo 3 (Google) leads on realistic physics and native audio. Kling 2.1 (Kuaishou) handles stylized character motion and expressive performance unusually well. Ray 2 sits between them, with strong cinematic camera moves and reliable image-to-video. None of the three is universally best; the right pick depends on the clip. For documentary-style realism, Veo. For stylized characters, Kling. For shot-driven, camera-led pieces and a clean image-to-video pipeline, Ray 2. Many editors keep two of the three models in rotation.

Writing one giant text-to-video prompt and rolling it five times. Text-to-video on any model is the highest-variance way to use it, and Ray 2 is no exception. The fix is mechanical: spend the first few credits generating the still you actually want (either Photon or any other image tool), then switch to image-to-video and write a short action plus a clear camera move. Variance drops, quality rises, and the credit bill stops being a horror show.