The short answer
Luma, made by Luma AI, is an AI video and image studio where Luma Agents plan and produce creative work from a brief. Its current video model, Ray3.2 (June 2026), makes clips up to 20 seconds at 1080p with up to 16 keyframes and native HDR. It also offers third-party models such as Kling 4.0 and Veo 3.1. There is no lasting free plan, only trial credits, and paid plans start at Plus.
| Made by | Luma AI, United States |
|---|---|
| Free plan | No lasting free plan; free trial credits to try it |
| Models | Ray3.2 for video, Uni-1 for images, plus Kling, Seedance, Veo and others |
| Best at | Controlled cinematic shots, reworking footage, ad variants at volume |
| Shows sources | Not applicable: it generates video and images |
| Team plans | Team and Enterprise, with shared credits and SSO |
| Your data | Privacy policy lists model training among uses of collected data, including uploads; enterprise terms set by contract |
What Luma is in late 2026
Luma is the creative app from Luma AI, a company based in Redwood City, California. If you last looked at it when it was called Dream Machine, it has changed a lot. Luma now lists Dream Machine and the older Ray2 model as deprecated. The current product is the Luma app, where Luma Agents take a brief and plan, generate and refine the video, images and audio for it. You still type prompts, but you can also ask the agent to handle a whole job, such as a set of product ads in three formats.
The engine underneath is Luma's own Ray family of video models. Ray3.2, released on 9 June 2026, is the current model. It makes clips up to 20 seconds long at 1080p, lets you place up to 16 keyframes inside one clip to control pacing and motion, and generates native HDR with 16-bit EXR export for colour grading. Ray3.14, released on 26 January 2026, brought native 1080p and was described by Luma as 3x cheaper and 4x faster than the version before it.
Luma is no longer only about its own models. The pricing page lists third-party video models such as Kling 4.0, Seedance 2.5, Veo 3.1 and MiniMax H3, image models such as GPT Image 2 and Nano Banana 2, and ElevenLabs voice, music and sound effect models. Kling 4.0 arrived on 23 October 2026.
What Luma is best at
Directed, cinematic shots
Ray3.2 is built for control rather than lucky accidents. Keyframes let you say what the frame looks like at the start, the middle and the end, and the model fills the motion in between. For a product reveal, a slow push into a building or a shot that has to land on a beat in a music track, this is far more reliable than describing everything in one prompt and hoping.
Reworking footage you already have
Modify Video V2 transforms existing footage up to 20 seconds at 1080p, so you can restyle a phone clip, change the setting or swap the look while keeping the movement. Reframe changes the aspect ratio, extends the frame or replaces the background after the take, which is useful when a horizontal shoot suddenly needs a 9:16 version for Reels.
Storyboards before you spend credits
Luma Scenes, launched on 11 August 2026, turns a brief and reference images into a storyboard of editable keyframes generated together, so the look stays consistent across scenes. You regenerate or comment on any frame, drag keyframes to set timing, and nothing renders until you approve. It renders with Ray 3.2 or Seedance 2 in 9:16 or 16:9. This is the feature that makes Luma practical for marketing teams, because the expensive step happens only once the plan is agreed.
Ad variants at volume
Luma Variants, live since 1 October 2026, takes one approved static ad and builds resized and translated versions while keeping the logo, headline and call to action in place.
Repeatable workflows
Luma Skills, added on 16 June 2026, saves a process that worked as a reusable workflow. You open the Skills menu at the bottom right of Luma Agents, give the Skill a name, description and plain-language instructions, then run it again on new assets with a slash command. Skills can be shared by link or downloaded as a zip file, so a team can agree on one way of making, say, product videos.
Where Luma falls short
- No lasting free plan. The pricing page starts at the paid Plus plan. Luma says free trial credits are included so you can try the platform, but it does not state how many.
- Credits are hard to predict. Every model costs a different number of credits, so a job that mixes Ray, Kling and Veo can burn through an allowance faster than you expect. Draft Mode, which uses fewer credits, helps while you explore.
- Short clips. A single generation is up to 20 seconds. Longer pieces are built from several shots, which is normal for AI video but still takes planning.
- Training on your uploads. On individual plans, the privacy policy lists model training among the uses of the information Luma collects, which includes your uploads, with no simple switch to stop it.
- Not a presenter tool. If you need a person talking to camera about your quarterly results, an avatar tool such as HeyGen or Synthesia fits better.
How to get better results from Luma
- Write the shot, not the story. Describe one shot at a time: subject, action, camera move, lens feel and light. "Slow dolly in on a ceramic mug on an oak table, morning window light, shallow depth of field" beats "a cosy coffee advert".
- Use keyframes for anything that must land. If the clip has to end on your logo or a product angle, set that as the last keyframe instead of asking for it in words.
- Storyboard first. For anything with more than two shots, start in Luma Scenes and approve the frames before rendering.
- Explore in Draft Mode. Try three or four directions cheaply, then render only the winner at full quality.
- Pick the model per shot. Use Ray3.2 when you need keyframes, HDR or reworking of footage, and try a third-party model when a shot needs something Ray struggles with. Compare two models on the same prompt before committing.
- Save what works as a Skill. Once a workflow gives you good results twice, turn it into a Skill so the next run starts from the same instructions.
Privacy settings and your data
Luma's privacy policy, updated on 20 April 2026, says it collects the prompts, images, videos and other files you upload, and it lists training and improving its AI models among the uses of the information it collects. It does not describe a training switch in the app. If you are in the EEA or the UK, the policy lets you object to processing by emailing hello@lumalabs.ai. Where Luma works for an enterprise customer, the enterprise agreement governs instead, so teams handling unreleased products or client footage should ask about training terms before uploading. A sensible rule: do not upload anything to an individual plan that you would not want inside a training set.
Which Luma plan suits whom
All paid plans include Luma's own and third-party models, and Plus already includes commercial use. The difference between Plus, Pro and Ultra is mainly the size of the monthly credit allowance and how much Luma Agents work you can do: Pro gives four times the agent usage and Ultra fifteen times. Start on Plus if you make a few campaigns a month, and move up only when you keep running out of credits. Team adds shared projects and credits, member management, usage analytics and SSO, which matters as soon as more than one person spends the budget. Enterprise adds custom terms, training and fine-tuning. Check the official pricing page for current prices.
Luma is a good choice for marketers, designers and small studios who want control over shots and many formats from one asset. It is not the right first stop if you just want to try AI video for free, where Pika or Google's tools are easier to start with. For a side-by-side view, see Runway vs Luma, the best AI for video and our Luma lessons.
Prompts to copy
Plan a product reveal with keyframes
I am making a 10-second product video for [product] in 9:16. Suggest three keyframes (start, middle, end) for Ray3.2, each as one sentence describing subject, camera position and light, then write the motion prompt that connects them. The last frame must show [logo or product angle].
You get: A keyframe plan you can paste straight into Luma
Storyboard a 30-second ad
Act as a creative director. Turn this brief into a six-scene storyboard for Luma Scenes: [paste brief]. For each scene give the shot type, camera move, what is on screen and the length in seconds, keeping the same character, colour palette and lighting throughout.
You get: A shot list ready to approve before you render
Restyle footage you already shot
Restyle this clip as [style, for example a 1970s film look with warm grain] while keeping the camera movement, the person's performance and the framing exactly the same. Change only the colour, texture and background details.
You get: A Modify Video prompt that keeps the motion intact
Turn one ad into a full set
Here is our approved ad: [upload]. Create versions for Instagram Stories (9:16), a LinkedIn feed post (1:1) and a website banner (16:9 widescreen), plus German and French versions of each. Keep the logo, headline and call to action in the same place and never change the product.
You get: A Variants brief that produces every size and language in one go
Write a Skill for repeat work
Write instructions for a Luma Skill called [name]. Input: one product photo. Steps: remove the background, place the product in [setting], create a 6-second slow orbit video in 9:16, and add a soft ambient sound. List the edge cases to handle, such as reflective or transparent products.
You get: Plain-language Skill instructions your team can reuse
Compare two models fairly
Generate the same shot with Ray3.2 and [Kling 4.0 or Seedance 2.5]: [describe shot]. Use identical length and aspect ratio, then list the differences in motion, faces, text on screen and lighting so I can pick one for the rest of the campaign.
You get: A like-for-like test before you commit credits