Luma AI is a creative AI platform where agents orchestrate Ray3.2, Uni-1 and third-party models such as Veo 3.1 and Kling across video, image and audio. Built for agencies and brand teams producing campaign assets at volume, with 1080p HDR output and ACES EXR export for professional post pipelines.
Used this tool? Rate it
Used this tool? Rate it
Luma AI is a creative AI platform built around a simple premise stated on its own homepage: Luma is your creative partner. You have the idea. Luma helps make it real. The product no longer presents itself as a single prompt box that turns sentences into clips. Its own metadata describes it as the creative AI platform for video generation and image creation, powered by the Ray and Uni model families and by agents that carry creative work end to end.
If you arrived here expecting Dream Machine, some orientation is needed. Dream Machine was the name that made Luma famous in 2024, and it still exists — but not as the front door. Visiting the old /dream-machine path now redirects to /app, which is titled Creative agents for creative professionals. The company's own API documentation refers to the "legacy Dream Machine video API" and offers a migration guide. On iOS the mobile app is still published under the Dream Machine name, so the brand persists on one surface while the web product has moved on. That split matters practically: if you search the App Store for "Luma Agents" you will not find the app.
The company behind it is Luma AI, Inc., based in Palo Alto, California. Its stated mission is to build unified general intelligence that can generate, understand, and operate in the physical world — a framing that positions video generation as a means rather than an end. CNBC, reporting independently in November 2025, described Luma as developing multimodal "world models" that learn from video, audio, and images rather than text alone, in order to simulate reality. In that same report CNBC confirmed a $900 million funding round led by Humain, an AI company owned by Saudi Arabia's Public Investment Fund, with participation from AMD's venture arm alongside existing investors Andreessen Horowitz, Amplify Partners, and Matrix Partners, valuing the company upwards of $4 billion. The same round pairs Luma with Humain to build a 2-gigawatt AI supercluster in Saudi Arabia, dubbed Project Halo, which the CEO described as one of the largest GPU deployments in the world — context worth knowing, because compute of that scale is what a company needs to keep shipping new video model generations at the pace Luma has been setting.
The word agent is heavily abused in AI marketing, so it is worth pinning down what it denotes in this product. Luma's own framing is that they are embedded in your workflow, advancing work end-to-end with shared context across teams. Three mechanics follow from that:
Luma's own FAQ gives the operating instruction plainly: your instructions should prioritise what you want, not how to do it — think of yourself as a director, not a micro-manager. That is a useful sentence to keep in mind, because it explains why prompts written like software commands tend to underperform here relative to prompts written like creative briefs.
The platform is not a wrapper around one model, and it is not exclusively self-built either. Luma develops Ray3.2, Uni-1, and Ray3.14 in house. Alongside them the platform orchestrates Nano Banana, Nano Banana Pro, GPT Image 2, Seedream, Veo 3.1, Kling 3.0, Kling Omni, Seedance 2.0, and three ElevenLabs audio models covering speech, sound effects, and music. This is unusual candour for a company that builds its own models — Luma publishes competing systems inside its own pricing table, and the Plus plan benefit is literally worded as access to "Luma and third-party image and video models."
The practical implication is that Luma competes on orchestration and workflow rather than on having the only good model. If Veo 3.1 is better for a particular shot, you can reach it without a second subscription.
Ray3.2 is the flagship video model, and what distinguishes it within the crowded field of AI video generation is control rather than raw novelty. Multi-Keyframe lets you set up to 16 keyframes inside a single clip, so you can direct what changes and what holds across the beats of a shot rather than accepting whatever motion the model invents. For anyone who has fought a video model that reinterprets the whole scene on every regeneration, this is the single most consequential capability on the page.
Resolution and length are stated concretely: 1080p output across every mode, with video-to-video work supported up to 20 seconds.
This is the capability that separates Luma from pure text-to-video tools. Modify Video V2 reshapes footage you already have — swap the wall, the world, the wardrobe, while the lighting holds and the performance survives, up to 20 seconds, all at 1080p. You bring real footage; the model changes the world around the performance instead of regenerating the performance.
The published transformation set covers eight named operations:
Supporting these are Improved Motion Transfer, Performance Tracking for body movement and posture, and Expressive Facial Performance for transferring facial expressions while holding character identity.
Most AI video tools stop at an MP4. Luma does not. Ray3.2 offers native 16-bit HDR generation — vivid, accurate, and gradeable from the first render rather than dependent on a post-hoc upconversion. More significant for professional pipelines is EXR export: 16-bit EXR frames for colour grading, compositing, and VFX in ACES2065-1 (AP0), comp-ready out of the box.
The company is explicit about why this matters, stating that native HDR generation and 16-bit EXR export exist so that AI work composites alongside live-action plates in your users' Resolve or Nuke. If you work in a facility where finishing happens in DaVinci Resolve or Nuke, this is the specification that determines whether generated material can enter your timeline at all or has to be treated as reference only.
Be aware that this quality carries a direct cost multiplier, covered under Pricing below.
On the image side, Uni-1 is positioned as brand intelligence at the model level, carrying style references, character references, and visual grounding into every asset. The API implementation reveals the architecture: one reasoning endpoint, one generation endpoint — the model thinks before it ships, so the first pass is usually the right one. Luma claims Uni-1.1 halves both price and latency relative to its predecessor. That claim is vendor-reported and has not been independently audited.
Four recent additions round out the production side:
Luma publishes a specific catalogue of the work it is built for, which is more informative than generic promises. The recurring pattern is volume with consistency — many variants of the same creative idea, held on-brand.
Infinite brand identity explorations — directing a full multi-asset campaign with cohesive visuals and variations ready to run everywhere — is the flagship scenario, and it is the one where the agent architecture earns its keep. Related scenarios include social media video ads for every channel, going from a text-to-video draft to short-form ads with hooks, captions, and per-platform formats; and smarter A/B testing variants for polished app store screenshots with platform-specific framing.
Global Campaigns is the version of this that most clearly saves money: generating market-specific variations of a shoot rather than reshooting per region.
Product visuals in all angles covers polished lifestyle and hero shots for e-commerce, social, and marketplaces in every required format. On-model photography with every angle produces realistic on-model imagery across diverse poses, body types, and styling contexts. Explore packaging & label mockups visualises packaging in photorealistic 3D, from shelf views to close-ups. Product Swap on the video side extends the same logic to motion: adapt products across campaigns without reshoots or manual editing.
Storyboards for every possible arc covers developing characters, story arcs, and episode boards, then bringing the pilot to life. Direct a book, game, or film trailer targets short cinematic pieces that set tone and build intrigue. This is the segment where HDR and EXR output matters most, because narrative work is finished in a colour pipeline rather than published straight to a feed.
Localize a video into multiple languages translates and localises video with natural voiceovers and synced visuals across languages. The ElevenLabs models in the roster are what make this practical inside a single tool rather than requiring a separate dubbing vendor.
Intelligence-driven lesson plans builds lesson plans complete with visuals and guides. Visualize academic content turns complex systems into labelled, easy-to-follow animated visualisations. Presentation-ready slide decks turns ideas into structured decks. Turn a podcast clip into a video converts podcast audio into shareable clips with captions, visuals, and B-roll. Transform data into infographics covers the static equivalent.
Luma's own recommended entry point is not the prompt bar. The Learning Center advises: jump in and start creating — follow guided workflows, then duplicate the board to remix, tweak, and make it yours. Boards are the organisational unit of the product; starting from a working example and modifying it teaches the system's conventions faster than starting empty.
Recall the official guidance: prioritise what you want, not how to do it; be a director, not a micro-manager. Practically, that means specifying subject, setting, mood, camera behaviour, and the intended deliverable format, then letting the agent choose the model and the sequence of operations. Over-specifying the mechanics tends to fight the orchestration layer.
The credit table makes the workflow discipline obvious. On Ray3.2, a 5-second SDR generation costs 20 credits in Draft, 50 at 540p, 100 at 720p, and 400 at 1080p. Drafting at low resolution costs one twentieth of a 1080p render. Explore composition and motion cheaply, then finish only the takes that work.
If usable source footage exists, Modify Video is almost always the better route than generating from scratch: the performance and timing are already correct, and you are asking the model to change the world rather than invent one. Set Motion and Structure appropriately, and note the official caveat that source frame rate affects output length.
For a shot that needs to hit specific moments — a product reveal, a turn, an expression change — place keyframes rather than describing the sequence in prose. With up to 16 per clip you have enough resolution to choreograph most single shots.
Decide SDR, HDR, or HDR+EXR before the final render, because the format multiplies the price threefold at the top end. Social delivery rarely justifies EXR. A spot going through grading always does.
Once a workflow produces reliable results, turn it into a Luma Skill. Build the creative workflow once, run it forever — this is where per-asset time cost actually collapses on recurring campaign work.
Advertising agencies and brand studios are the clearest fit, and the customer roster the company displays reflects that: Serviceplan, Dentsu, Humain, Aperture, Publicis Groupe, Boundless, Mazda, and Uncharted. These are vendor-selected references rather than independently verified case studies, but the pattern of who is being sold to is unambiguous. The named testimonials come from Aperture CEO Hannah Paravez, Boundless Co-Chief Creative Officer Fran Luckin, The Wonder Project founder Jon Erwin, and Publicis Groupe Middle East & Turkey CEO Bassel Kakish.
In-house brand and content teams producing recurring campaign assets are where creative agents pay for themselves, and they benefit most from Skills and shared context, where the same workflow runs many times with different inputs.
Post-production and VFX facilities are served specifically by the EXR and ACES2065-1 output path. If your finishing happens in Resolve or Nuke, Luma is one of the few generative tools that outputs into that pipeline rather than beside it.
Product and developer teams building generation into their own applications are addressed by the API, with dollar-denominated pricing and a Scale tier carrying an SLA.
Educators and course creators get lesson planning and academic visualisation as first-class scenarios, and Luma runs a dedicated education program.
Who it fits less well: individual hobbyists and casual experimenters. There is no free plan, entry is $30 per month, and the credit economics reward volume. Someone making a handful of clips a month will find the per-clip cost hard to justify against tools with free tiers.
Web is the primary environment. All paths converge on app.lumalabs.ai as the single sign-in and creation entry point, and the browser is where boards, agents, and team workspaces live.
iOS has a native app published as Luma Dream Machine, seller Luma AI, Inc., free to download with subscription and in-app purchases separate. The current version at the time of writing is 2.0.8, released 17 April 2026, requiring iOS 17.0 or later. Apple classifies it under Photo & Video with a secondary Graphics & Design category, rated 12+.
API access is a separate product line with its own dollar pricing, covering Ray3.2 for video and Uni-1.1 for images.
Professional post pipelines are supported indirectly but concretely through ACES2065-1 EXR output designed to composite alongside live-action plates in Resolve or Nuke.
Android availability could not be confirmed from primary sources at the time of writing, so no claim is made in either direction. (To be verified)
This is the first thing to establish, because a great deal of third-party content states otherwise. Luma's pricing page lists no free tier. The cheapest individual entry point is the paid Plus plan.
The terms do define Trials — alpha, beta, or early access offerings — but with sharp limits: not to exceed 14 days unless otherwise agreed in writing, and critically, any output generated during a trial may only be used for internal evaluation and testing of the trial and for no other purpose. Luma provides no warranty, indemnity, or support for trials, and caps liability for them at US $50. A trial is therefore not a way to produce publishable work.
| Plan | Monthly | Yearly (per month) | Credits | Key inclusions |
|---|---|---|---|---|
| Plus | $30 | $25 ($300/yr) | 10,000 | Luma and third-party image and video models; edit access for guest collaborators; commercial use |
| Pro | $90 | $75 ($900/yr) | 40,000 | Everything in Plus, plus 4x usage with the Luma Agents |
| Ultra | $300 | $250 ($3,000/yr) | 150,000 | Everything in Pro, plus 15x usage with the Luma Agents |
Annual billing is advertised as saving up to 20%.
Note that commercial use appears as a listed Plus benefit, which aligns with the terms treating commercial rights as granted by subscription rather than as automatic. See Limitations below.
Team (contact sales) adds member management, projects and team-wide sharing, usage analytics, shared team credits, and SSO. Enterprise adds enterprise commitments, dedicated education and training, and custom fine-tuning.
Credits are not a flat currency — consumption varies by model, action, resolution, duration, and output format. Selected Ray3.2 video rates:
| Action | Draft | 540p | 720p | 1080p |
|---|---|---|---|---|
| T2V / I2V, 5 sec | 20 | 50 | 100 | 400 |
| T2V / I2V, 10 sec | 60 | 150 | 300 | 1,200 |
| Reframe, per sec | 10 | 20 | 40 | 120 |
Output format multiplies the base: SDR is base pricing, HDR is 2× SDR credits, and HDR + EXR is 3× SDR credits. A 1080p 10-second HDR+EXR clip therefore costs 3,600 credits — more than a third of the entire monthly Plus allocation for a single shot.
Cross-model rates vary by more than an order of magnitude. Veo 3.1 with audio runs 280 credits per second at both 720p and 1080p; Seedance 2.0 Mini at 480p costs 26 credits per second. Image generation ranges from 1 credit for a Seedream 1K image to 255 credits for GPT Image 2 at High 4K, with Uni-1 at a flat 30 credits. Audio is metered separately: ElevenLabs v3 text-to-speech at 21 credits per 1,000 characters, sound effects at 25 credits per minute, music at 98 credits per minute.
The API is priced in dollars rather than credits. Ray3.2 text-to-video and image-to-video for a 5-second clip: $0.15 at 540p, $0.30 at 720p, $1.20 at 1080p. Prices shown are for SDR; HDR is 2× and HDR+EXR is 3×.
Two commercial tiers exist. Build is pay-per-use with no minimum commitment, but the company states plainly that rate limits apply and there is no latency SLA. Scale provides dedicated capacity with guaranteed throughput and latency, includes SLA, moderation, and prompt enhancement, and carries a no-train guarantee. Scale requires a minimum of 4 units, where 1 unit equals 1 request per minute on Base or 0.4 RPM on Max.
Subscriptions automatically renew for successive terms at then-current rates unless cancelled before the renewal date. All fees are final and non-refundable unless otherwise determined by Luma. Fee disputes must be raised in writing within 60 days of billing. Payments are processed by Stripe, not by Luma directly. Given the refund complaints discussed under Limitations, these clauses are the operative context.
Because Luma orchestrates competitor models internally, "alternatives" splits into two questions: alternative models, and alternative platforms.
Runway is the closest platform-level competitor, with a comparable professional editing orientation and a longer history in film production workflows.
Pika targets shorter-form social content with a lighter interface and generally lower price point, appealing to creators rather than agencies.
Google Veo and Kling are available inside Luma, so choosing them directly is mainly a question of whether you want the orchestration layer or direct model access at potentially lower cost.
OpenAI Sora competes at the model level. Luma's CEO has publicly claimed Ray3 benchmarks higher than Sora 2 and roughly level with Veo 3 — a vendor claim reported by CNBC but not independently audited.
Adobe Firefly appeals where Creative Cloud integration and commercially safe training data are decisive procurement factors.
Where Luma is genuinely differentiated: the ACES EXR output path, 16-keyframe control, and multi-model orchestration under one subscription. Where competitors may win: free tiers for evaluation, and single-model simplicity for teams that do not need orchestration.
This is the most important thing a prospective buyer should understand, and it is not visible from the marketing site.
On the App Store, Luma Dream Machine holds 4.5 out of 5 across 1,780 ratings — a large sample and a strong score.
On Trustpilot, lumalabs.ai holds 1.5 out of 5 across 62 reviews. Two qualifiers are essential for reading that number honestly. First, Trustpilot itself flags that the profile is unclaimed and that the company has no history of inviting customers to review, so the sample may not be representative — self-selected reviewers skew negative. Second, the entire category scores low there: on the same page, Runway shows 1.1 (307 reviews), Kling 1.3 (365 reviews), and Pika 1.7 (66 reviews). Luma is not an outlier within its category.
What is useful is not the score but the recurring themes, which cluster tightly:
Read against the terms — all fees final and non-refundable unless Luma decides otherwise — the third theme is less about individual support failures and more about a policy that leaves little room for recovery. Budget accordingly.
Luma's own API FAQ is unusually forthcoming about defects, and reading it is worthwhile before committing to an integration. Documented issues include HDR requests falling back to Ray 3 instead of Ray 3.2; Extend failing with a generation error; the Reframe panel using only Ray 2 with no engine selector; imported videos failing on Modify with a download error; V2V Modify ignoring source video motion and generating from the prompt alone; and multiple videos being generated from a single click, causing duplicate charges.
The terms restrict using the services or output for commercial purposes unless permitted under Section 4.9, and separately prohibit distributing or selling output through any marketplace not operated by Luma without a subscription permitting commercial use. In practice a paid plan carries commercial rights — Plus lists it explicitly — but free-form assumptions about commercial rights are misplaced. Trial output is restricted to internal evaluation regardless of tier.
The privacy policy lists among its purposes: train, develop, and improve the artificial intelligence, machine learning, and models that we use to support our Services. Set against the API Scale tier's explicit no-train guarantee, the picture is clear — freedom from training use is an enterprise feature, not a default. For agencies handling client-confidential material before announcement, this is a procurement-level consideration.
The terms state that the customer owns and retains all right, title, and interest in the output, with Luma assigning its own interest. But the assignment does not include any right, title, or interest in output of any third party — relevant given that similar prompts can produce similar results for different users.
If you use output that identifies or resembles a person, you must publicly identify that the output was AI generated. You must not misrepresent the source of output or claim it is human-generated. Uploaded material containing people requires that you own or are fully authorised to grant rights in it.
The terms prohibit using the services or output, directly or indirectly, to create, test, train, or otherwise develop any AI or ML models, systems, architecture, weights, or related technology. Synthetic training data generation is off the table.
Luma may modify or discontinue all or any part of the services at any time, temporarily or permanently, without notice — with a commercially reasonable effort to give 30 days' notice for deprecation of material functionality — and disclaims liability for the consequences. The Dream Machine to Agents transition demonstrates this is not hypothetical.
Ray3's positioning as the first reasoning video model and the first to generate 16-bit HDR comes from the company. The comparative benchmark claim against Sora 2 and Veo 3 is the CEO's, reported by CNBC. At the time of writing, no independent hands-on technical evaluation from an established technology publication was located. CNBC also noted that Dream Machine drew accusations of copying intellectual property earlier that year, with the company responding that safeguards had been installed.
Users must be at least 18, or between 13 and 18 with parental or guardian consent. Users under 13 are not authorised, and a COPPA notice applies.
No. The pricing page lists no free plan; the cheapest individual entry point is Plus at $30 per month, or $25 per month billed annually. Trials of alpha, beta, or early access features exist but are capped at 14 days and their output may only be used for internal evaluation and testing, not for publishable work. Some third-party articles state that a free tier exists — that does not match the current pricing page.
Yes on a paid plan. Commercial use is listed as an explicit benefit of the Plus plan, and the terms permit commercial use where allowed under Section 4.9. You own the output, with Luma assigning its rights to you. Two caveats: the assignment excludes rights in any third party's output, and trial output is restricted to internal evaluation regardless of plan.
Ray3.2 outputs 1080p across every mode, with 540p, 720p, and a Draft tier below it. Video-to-video work is supported up to 20 seconds. Other models in the roster reach higher — Seedance 2.0 and Kling Omni offer 4K options — at correspondingly higher credit rates.
Credits are consumed per generation based on model, action, resolution, duration, and output format. A 5-second 1080p Ray3.2 clip costs 400 credits in SDR; the same clip in HDR+EXR costs 1,200, since HDR is 2× and HDR+EXR is 3×. With 10,000 credits on Plus, that is roughly 25 finished 1080p SDR clips before accounting for iteration — and iteration is where budgets actually go. Draft mode at 20 credits per 5 seconds exists precisely for this reason.
The name persists on iOS, where the app is published as Luma Dream Machine. On the web it does not: /dream-machine redirects to /app, presented as Luma Agents, and the API documentation describes the Dream Machine video API as legacy with a migration guide. The underlying video generation capability continues under the Ray model line.
Yes, and this is one of Luma's stronger capabilities. Modify Video V2 reshapes existing footage — changing environments, wardrobe, lighting, or products — while preserving the original performance, at up to 20 seconds and 1080p. Related tools include Motion Transfer, Camera Motion Transfer, Relighting, and Reframe for aspect ratio changes. Note the documented caveat that V2V Modify can sometimes ignore source motion and generate from the prompt alone.
On consumer subscriptions, the privacy policy lists training, developing, and improving Luma's AI and machine learning models among its stated uses of collected information. The API's Scale tier explicitly includes a no-train guarantee. If training exclusion is a requirement, that points to the enterprise API tier rather than a standard subscription.
For finishing pipelines, yes — more so than most generative tools. Ray3.2 generates native 16-bit HDR and exports 16-bit EXR frames in ACES2065-1 (AP0), described by Luma as comp-ready and designed to composite alongside live-action plates in Resolve or Nuke. Budget for the 3× credit multiplier that HDR+EXR output carries.
It depends heavily on where you look. The iOS app holds 4.5 out of 5 across 1,780 ratings. Trustpilot shows 1.5 out of 5 across 62 reviews, though that profile is unclaimed with no review invitations, and the whole category scores low there — Runway 1.1, Kling 1.3, Pika 1.7. The consistent critical themes across reviews are day-to-day variance in output quality, credits consumed by failed attempts, and difficulty obtaining refunds. Note that the terms make fees non-refundable by default.
Yes, covering Ray3.2 for video and Uni-1.1 for images. Ray3.2 5-second SDR clips are priced at $0.15 for 540p, $0.30 for 720p, and $1.20 for 1080p, with HDR at 2× and HDR+EXR at 3×. The Build tier is pay-per-use with no minimum commitment but explicitly no latency SLA. The Scale tier provides dedicated capacity with guaranteed throughput, an SLA, moderation, and a no-train guarantee, with a minimum of 4 units where 1 unit equals 1 request per minute on Base.
