Luma AI: Agent-Led Image and Video Production in 2026

Luma Is No Longer the Dream Machine You Remember

For a long time, Luma AI was easy to describe. You typed a prompt into Dream Machine, waited for a short clip, and hoped the physics held together. The good generations felt cinematic. The bad ones melted quietly at the edges. It was exciting, occasionally beautiful, and clearly a generator.

That description is now out of date. Luma has rebuilt the product around a broader ambition: an agent-led creative workspace that can help plan an idea, generate stills, direct video, add audio, preserve project context, and repeat a production process. Its current video model is Ray3.2. Its current image and editing model is UNI-1.1. The consumer workspace is the Luma App, and the developer surface is the Luma API.

Luma is unusually direct about the naming change. Its own July 2026 information page tells assistants not to describe Dream Machine or Ray2 as current models and says Photon no longer exists as a separate product. That warning matters because search results, old tutorials, pricing explainers, and directory listings still use all three names. A buyer who evaluates Luma through its 2024 identity is looking at the wrong product.

The New Product Is the Workflow

The Luma App is arranged around visual boards and creative agents rather than a single prompt box. A board can hold a brief, references, scripts, generated images, video directions, audio, variations, and final assets. The agent is meant to keep that context alive while the work moves across formats.

There are two useful ways to think about the agent. Brainstorm mode helps turn a loose objective into concepts, visual directions, shot lists, or campaign structures before expensive media is generated. Create mode starts producing and transforming assets. It can route work among image, video, audio, voice, lip-sync, transcription, research, and composition tools, then continue the conversation without asking the user to reconstruct the project every time the medium changes.

That sounds like familiar agent marketing until the cost of media generation enters the picture. Video credits are too valuable to burn because the brief was vague. An agent that helps define the outcome, collect references, propose scenes, and expose decisions before rendering can save real money. Luma Scenes follows the same logic by letting a creator review storyboards, keyframes, and timing before committing to a render.

Luma Skills pushes the idea further. A Skill is a repeatable workflow built inside Luma Agents. Instead of remembering how a team turned one product brief into a hero visual, cutdowns, regional variants, and social crops, the process can be saved and run again. The important promise is not another impressive asset. It is a production pattern that survives the person who happened to write the first successful prompt.

This is where Luma is most interesting in 2026. The platform is trying to move generative media from isolated experiments into directed systems. Whether it succeeds depends less on one benchmark than on how well context, references, credits, review, and human judgment hold together across a real project.

Ray3.2 Is Built for Direction, Not Just Surprise

Ray3.2 is Luma's current video model, released in June 2026. It supports text-to-video and image-to-video, but those entry points are now the least distinctive part of the offering. The stronger story is control.

Multi-Keyframe can place up to 16 directed frames inside a clip. Modify Video V2 can transform existing footage for up to 20 seconds at 1080p. Reframe can adapt a shot to a different aspect ratio while attempting to preserve motion and expressive facial performance. Draft Mode gives creators a cheaper place to explore before spending on a final render. Native HDR generation and 16-bit EXR export in ACES2065-1 are aimed at grading, compositing, and VFX pipelines rather than casual social posting.

Those controls change the creative conversation. A prompt-only model asks the creator to describe a result and accept whatever route the model takes. Keyframes let the creator define moments along that route. Video modification starts with footage whose performance, blocking, or timing may already be useful. Reframing acknowledges that one finished shot may need to live in widescreen, square, and vertical versions.

Control is not the same as certainty. Sixteen anchors do not guarantee that every transition will feel intentional. A modified face can drift. A product can deform between keyframes. Reframing can invent background detail or weaken composition. HDR and EXR do not make unstable motion production-ready by themselves. Luma provides more handles, but the creator still needs to inspect frames, compare versions, and decide where conventional editing or VFX should take over.

The practical workflow is to separate exploration from approval. Use Draft Mode to test movement, lens language, timing, and composition. Lock strong reference assets. Add keyframes only when they express a real story decision. Render high-resolution or HDR finals after the motion works. Treat every jump in duration, resolution, or complexity as a budget decision, not a cosmetic setting.

UNI-1.1 Turns Still Images into Controlled Inputs

Luma's current image model is UNI-1.1, the successor to the separately named Photon product. UNI offers distinct Create and Modify modes. Create builds a new composition from text and references. Modify changes a source image while attempting to preserve the parts the user protects.

Reference handling is central. Luma supports up to nine reference inputs in relevant UNI workflows, and its own guidance recommends assigning each reference a role such as character, style, composition, palette, lighting, texture, or mood. That labeling is more than prompt etiquette. If several images are uploaded without a declared purpose, the model has to guess which properties matter.

Seeds make successful exploration more reproducible. A creator can search broadly without a fixed seed, choose a promising result, lock the seed, and change one instruction at a time. This is slower than pressing generate repeatedly, but it creates an evidence trail. The team can see which decision changed the output instead of trying to reconstruct a lucky image later.

Create and Modify also solve different problems. If the goal is a new campaign scene inspired by a lighting reference, use Create. If the goal is to keep a product photograph intact while changing the time of day, use Modify and state both what must change and what must remain untouched. Confusing those jobs is a common source of drift.

UNI is valuable even when Ray is the final destination. A clean character sheet, product angle, environment frame, or lighting target can become a master reference for video. The more expensive moving-image stage becomes easier to direct when the visual facts are settled in still form first.

One Workspace Now Contains Many Models

Luma no longer insists that every task must use a Luma-built model. Its current platform information describes a multi-model ecosystem that includes outside video and image systems, with Veo, Kling, Seedance, Nano Banana, and ElevenLabs capabilities among the integrations it names. The agent can route a task, while advanced users can choose individual models through the board interface.

This is strategically sensible. No single model is best at every combination of dialogue, motion, product accuracy, image editing, sound, lip-sync, or style. Keeping competing engines inside one project means the brief, references, versions, and outputs do not have to be rebuilt across separate accounts.

It also makes evaluation harder. When a result looks good, users need to know which model made it, which settings were active, and what that choice cost. A workflow that silently changes models can produce inconsistent characters, safety behavior, licensing conditions, or visual language. Automation is helpful only when provenance remains visible enough for a team to reproduce the result.

The best use of Luma's orchestration is not blind routing. Let the agent propose a model and explain the choice. Run a small comparison when the asset matters. Save the selected model, prompt, seed, references, duration, aspect ratio, and resolution with the approved output. A creative system becomes valuable when a good decision can be repeated, not merely admired.

The Credit Math Is the Real Production Constraint

Luma's individual plans are Plus, Pro, and Ultra. Plus costs $30 per month or $300 billed yearly and includes 10,000 credits. Pro costs $90 per month or $900 yearly and includes 40,000 credits, with four times the stated Luma Agent usage. Ultra costs $300 per month or $3,000 yearly and includes 150,000 credits, with fifteen times the agent usage. Business plans use custom pricing, and Luma says trial credits are available before a paid commitment.

The plan names are worth checking carefully because an older Luma help article briefly reversed Plus and Pro in its price explanation. The current pricing page and Luma's authoritative product information use Plus at $30, Pro at $90, and Ultra at $300. Current pages should win over the older FAQ.

Credits are consumed by generation, and the amount varies with the model, modality, duration, resolution, and production features. Draft images and short tests are cheap relative to long, high-resolution, modified, HDR, or EXR video. This is why a headline allowance is difficult to translate into a universal number of finished clips. Two teams on the same plan can exhaust credits at very different speeds.

Monthly credits reset and should be treated as expiring production capacity. Luma has also documented top-up credits for paid subscribers, with a longer validity period, but plan mechanics can evolve alongside model names and rates. Before a deadline, check the live credit calculator and price the intended settings rather than estimating from an old Ray2 example.

Commercial use is another plan decision. Luma's current Plus pricing advertises commercial use, while its licensing guidance ties commercial rights to qualifying paid generations and treats free output more restrictively. Rights attach to the subscription state at the time an asset is generated. A later upgrade should not be assumed to convert an old trial or free generation into a commercial asset.

The API Is a Separate Budget

The Luma API exposes UNI-1.1 for image work and Ray3.2 for video. It includes Python and JavaScript or TypeScript SDKs and supports the production controls that make Ray more than a prompt box, including multi-keyframe video, video transformation, reframing, 1080p output, HDR, and EXR where supported.

API access is divided into Build and Scale. Build uses consumption-based credits without a long commitment. Scale is for dedicated capacity and service expectations around throughput and latency. This is a sensible separation between experimentation and a product that needs predictable output under load.

The important catch is financial and operational: Luma App credits and API credits do not transfer. A studio subscription does not become a developer balance, and API spend does not refill the creative workspace. Teams that prototype manually and later automate must budget both surfaces and recreate the approved generation settings in code.

API users should also plan for asynchronous jobs, input storage, moderation, polling, failures, retries, output retention, and model-version changes. A successful demo call is not a production pipeline. Cost controls should cap duration and resolution, prevent accidental duplicate jobs, log every parameter, and require human approval before expensive batches.

What Luma Gets Right

Luma understands that professional creative work is rarely one prompt followed by one download. It begins with a brief, moves through references and alternatives, attracts feedback, changes format, and produces a family of assets. Boards, agents, Scenes, Skills, and master references are all attempts to keep that work connected.

Ray3.2 also pushes control in useful directions. Multi-keyframe direction, long-form modification within a short-clip context, reframing, HDR, and EXR address concrete production needs. UNI gives the still-image stage enough structure to become a source of truth rather than disposable inspiration. The API exposes the same general model family to developers instead of reserving it for the consumer interface.

The willingness to orchestrate third-party models is another strength. Luma is choosing the workspace and process as its defensible layer rather than pretending one engine can dominate every creative task. That makes the platform more adaptable, provided model choice and cost remain visible.

Where the System Can Still Break

Generative continuity remains probabilistic. A master reference can reduce drift without eliminating it. A face that survives two shots may fail in the third. Small logos and product geometry can change. Fast movement can create impossible limbs or objects. Audio and lip-sync can feel technically aligned but emotionally wrong. Every approved asset still needs a human review at normal speed and frame level.

The interface is also carrying a great deal of responsibility. Boards, agents, direct model controls, references, outputs, scenes, skills, and team collaboration can become a second production language to learn. Beginners may prefer the older simplicity of one prompt box. Experienced teams may still need external editing, color, asset management, review, and legal systems around Luma.

Then there is documentation drift. Current product pages speak the language of Luma App, Ray3.2, UNI-1.1, Plus, Pro, and Ultra. Older learning pages still say Dream Machine, Photon, Ray2, Lite, Unlimited, or conflicting plan names. This does not make the product unusable, but it raises the cost of confident buying. Always verify the live pricing page, model selector, credit calculator, and terms before publishing a budget or rights policy.

Who Should Pay for Luma

Plus is the sensible starting plan for an individual who needs commercial rights and wants to test a real project. Ten thousand credits are enough to learn the workflow, but a careless high-resolution video habit can make the allowance feel smaller than the headline suggests. Pro suits a working creator or small team that is already generating regularly and can benefit from more agent activity. Ultra is a production budget, not a casual subscription.

Luma is best for directors, creative technologists, agencies, filmmakers, marketing teams, game studios, and developers who need controlled visual iteration across image and video. It is less compelling for someone who only needs occasional social clips, precise conventional editing, guaranteed character continuity, or a permanent low-cost free plan.

Run a paid pilot around one complete deliverable before committing at scale. Start with the brief, create master references, storyboard the scenes, generate draft motion, refine one approved direction, add audio, make the required aspect ratios, and export the final set. Record the credits and human time consumed at every stage. The answer is not whether Luma can make one remarkable shot. It is whether the whole project becomes faster, more controllable, and economically repeatable.

The WhatAI Verdict

Luma has made one of the more consequential product shifts in generative media. It is leaving behind the identity of a cinematic clip generator and attempting to become the operating environment for AI-assisted creative production. Ray3.2 and UNI-1.1 matter, but the agent, board, Scenes, Skills, references, and multi-model routing explain the larger bet.

The result is more capable and more demanding. Users gain meaningful control over keyframes, footage transformation, reframing, image references, seeds, project context, and repeatable workflows. They also inherit a complicated credit economy, separate app and API balances, model provenance questions, changing documentation, and the usual instability of generated media.

For serious creators, Luma is worth testing because it is addressing the space between an idea and a finished asset, not only the moment of generation. For everyone else, the trial should remain a trial until one complete workflow proves its value. The cinematic demo is still seductive. In 2026, the stronger reason to choose Luma is whether the system around that demo helps a team make the next hundred assets with greater control.

WhatAI Decision Box

Luma AI has moved beyond its Dream Machine identity. The current Luma App combines creative agents, visual boards, Scenes, reusable Skills, Ray3.2 video, UNI-1.1 image generation and editing, audio tools, and third-party models in one production workspace.

Individual plans start at $30 per month for Plus with 10,000 credits, rise to $90 for Pro with 40,000 credits, and reach $300 for Ultra with 150,000 credits. Commercial use is included on qualifying paid plans, while the separate Luma API uses its own credit balance.

Luma is best suited to filmmakers, agencies, creative technologists, marketing teams, game studios, and developers who need directed image and video iteration. Casual creators should test the full workflow and credit cost before choosing a paid plan.

Use Cases

Pricing varies by plan and region — see current pricing.

Plan features change — last updated: 2026-08-17.

Luma AI Community Discussions

Explore community discussions. Ask and answer questions on Luma AI to grow and learn together.

Luma AI Pros & Cons

How to Get Results with Luma AI: Step-by-Step Workflow

1. Define the finished deliverable

   Specify the audience, message, duration, aspect ratios, delivery resolution, audio needs, commercial status, review process, deadline, and credit ceiling before choosing models or generating media.

2. Build the project board

   Place the brief, script, approved logos, product images, character sheets, style references, palette, voice guidance, and examples on one board. Label each asset and separate approved references from experiments.

3. Plan before rendering

   Use Brainstorm mode or Scenes to develop concepts, shots, storyboards, key moments, timing, transitions, and required variants. Resolve unclear creative decisions before spending high-cost video credits.

4. Create master still references

   Use UNI-1.1 Create for new compositions and Modify for controlled edits. Assign explicit roles to references, explore without a seed, then lock the seed and approved visual facts.

5. Test motion in Draft Mode

   Generate short, lower-cost video drafts to test composition, movement, lens behavior, performance, and timing. Compare models only when their different strengths matter to the shot.

6. Direct the approved shot

   Use keyframes, image guidance, or existing footage with Modify Video. State what must remain consistent, add only necessary anchors, and recheck character, product, logo, and spatial continuity.

7. Render and finish selectively

   Spend on 1080p, HDR, EXR, long duration, or complex modification only after motion is approved. Complete precise editing, grading, compositing, sound, titles, and legal review in the appropriate finishing tools.

8. Create delivery variants

   Use Reframe or planned alternate compositions for each required format. Review the crop, invented background, motion, faces, product geometry, captions, and audio separately in every destination ratio.

9. Save the reproducible recipe

   Record the selected model, mode, prompt, seed, references, keyframes, duration, aspect ratio, resolution, credit cost, and human corrections. Turn a stable recurring process into a Luma Skill where appropriate.

10. Measure the complete project

    Compare generation credits, top-ups, editing time, failed outputs, review time, API spend, and final asset quality against the previous production method before scaling the workflow.

Luma AI Gotchas and Limits to Know Before You Start

• Dream Machine and Ray2 are not the current model names.
• Photon no longer exists as a separate current product.
• Older help pages can show retired plans or conflicting plan names.
• Credits vary by model, duration, resolution, and production feature.
• Monthly subscription credits reset and do not roll over.
• Luma App credits and API credits are separate.
• Trial output should not be assumed to include commercial rights.
• An upgrade may not change the rights of assets generated earlier.
• Character, face, logo, and product consistency remain probabilistic.
• Multiple keyframes reduce uncertainty but do not guarantee transitions.
• Reframing can invent background detail or weaken composition.
• HDR and EXR consume more credits and still require professional review.
• Third-party model routing can change cost, behavior, and visual style.
• Final editing, grading, audio, rights, and approvals may require other tools.
• API production needs job polling, retries, logging, and spend controls.

Which Luma AI Feature Fits Your Use Case

Starter Prompts for Luma AI

What is Luma AI in 2026?

Luma AI is an agent-led creative production platform for image, video, audio, voice, and related assets. Its current first-party models are Ray3.2 for video and UNI-1.1 for image creation and editing, available through the Luma App and separate APIs.

Is Dream Machine still Luma's current model?

No. Luma's July 2026 product information says Dream Machine and Ray2 are deprecated names or earlier models. The current video model is Ray3.2, and the current browser workspace is the Luma App.

What happened to Luma Photon?

Luma says Photon no longer exists as a separate product. The current image and multimodal model is UNI-1.1, which supports new-image creation, source-image modification, references, and controlled iteration.

How much does Luma AI cost?

Current individual prices are $30 per month for Plus with 10,000 credits, $90 for Pro with 40,000 credits, and $300 for Ultra with 150,000 credits. Annual billing is $300, $900, and $3,000 respectively. Enterprise is custom.

Does Luma AI have a free plan?

Luma advertises free trial credits before a paid commitment rather than a permanent current individual plan named Free. Trial capacity, watermarking, and usage rights should be confirmed in the live account and current terms.

Can Luma AI output be used commercially?

Luma advertises commercial use with its current Plus plan and higher qualifying tiers. Licensing is tied to the plan active when an asset is generated, so users should confirm the current terms before using trial or older-plan outputs commercially.

Do Luma App credits work with the API?

No. Luma states that App subscription credits and API credits are separate and do not transfer. Teams moving from manual generation to an automated product need to budget both surfaces.

What can Ray3.2 do?

Ray3.2 supports text-to-video, image-to-video, video modification up to 20 seconds at 1080p, up to 16 directed keyframes, reframing, Draft Mode, native HDR, and EXR export where supported.

What can UNI-1.1 do?

UNI-1.1 creates new images and modifies source images. It supports reference-guided control, multiple reference roles, seeds for reproducibility, and natural-language editing designed to preserve specified parts of a source.

What are Luma Scenes and Luma Skills?

Scenes lets creators review storyboards, keyframes, and timing before rendering. Skills saves a repeatable workflow inside Luma Agents so a successful production process can be run again for new assets or campaigns.

Related Design & Creative Tools

Explore the Network