Luma Dream Machine

From Dream Machine to the current Luma environment

Dream Machine is the name of Luma’s original application, which the company retains as a legacy offering. Current development centres on Luma App, a workspace with creative agents, its own Ray video and UNI image models, and external models.

This profile retains the name through which many users discovered the service while distinguishing the historical product from the current offering. A 2024 Dream Machine guide does not necessarily describe today’s interface or plans. The central idea is preparing assets, generating alternatives, editing and organising production in a shared space; the specific technology depends on the selected operation and should be recorded when comparing results.

Reference documentation: About Luma · Luma current application.

3D origins and audiovisual evolution

Amit Jain and Alex Yu founded Luma in 2021. The company began with scene capture and neural rendering; Genie added 3D-object generation with meshes and materials. Dream Machine arrived in June 2024 and was redesigned in November. Ray2 was announced in January 2025, followed by Reframe and Modify Video.

This trajectory connects two different problems: representing a scene and generating how it might look or move. A cinematic image is not a measured geometric model. That distinction matters when depicting a workplace: the appearance of depth does not establish accurate dimensions, distances or physical behaviour, even when the camera movement seems convincing.

Reference documentation: About Luma · Genie1.0 and SeriesB · Product changelog.

Boards, agents and shared context

Luma Agents organises work on visual boards containing references, versions and assets. The agent interprets instructions and coordinates image, video and audio models. Its guide describes real-time collaboration and asset search within project context.

A useful board distinguishes approved sources, explorations and deliveries. For a campaign, place the message and constraints near the references illustrating them. When an instruction changes, record the affected pieces. Visual proximity helps comparison but does not itself constitute approval. The team needs to identify who decides content, which version is current and which alternatives have been rejected. A generated asset should not become the approved reference merely because it is prominent on the canvas.

Reference documentation: Welcome to Luma Agents.

Ray: generation, references and framing

The current Ray3.2 API distinguishes generation from text or images, editing existing video and reframing. Reference frames guide compositions and moments. Reframe adapts the framing, whereas generating a transition asks the model to produce content between visual states.

These operations are not interchangeable. A change of perspective may need only reframing; narrating an action requires deciding which movement to depict. If an area lies outside the original image, expanding the view may invent its contents. In safety material, keep labels and measurements requiring exact agreement with approved information outside generative processing. A visually continuous extension is still newly created material rather than recovered footage.

Reference documentation: Ray3.2 current product · Ray3.2 generation API.

Modify: transforming a recorded performance

Ray3.2 Modify starts from video and preserves its duration, up to twenty seconds. Controls separate movement, structure and character attributes, with keyframes guiding specific moments. The guide acknowledges that subtle facial details may not transfer completely.

The original recording becomes a performance reference. Changing clothing or surroundings can help explore a proposal, but object contact and posture also require review. Compare original and transformed footage side by side and pause at critical moments. If the explanation depends on how a tool is held, resemblance of the overall character is insufficient. Preserve the unmodified recording so the team can distinguish original performance from reconstructed details.

Reference documentation: Ray3.2 video-to-video guide.

Scenes: organising a sequence before rendering

Luma Scenes turns a brief into an editable storyboard. It supports revising individual frames, adjusting timing and producing video, separate clips or images. The documented Ray path does not generate audio; the rendering model is chosen separately from planning.

For awareness material, assign each shot a function: establish context, show a decision and finish with a question. Check whether information is missing between moments. A spectacular frame may interrupt the story’s logic. When a correction affects several shots, check all affected frames: the guide states there is no universal automatic propagation. Retain the approved sequence as the reference for later language versions and associated supporting materials.

Reference documentation: Luma Scenes.

UNI, image editing and layers

UNI combines multimodal understanding and image generation. Layers can separate an image into editable elements with transparency, move them, edit text and reconstruct the composition. Individual elements can also return to the board for use in other assets.

Reconstructed separation does not necessarily recover the original design file. An extracted logo may contain alterations, and a previously hidden background may be invented. For a multilingual campaign, keeping illustration and text separate helps replace the language without regenerating the entire piece. Compare final text against approved wording and retain an editable copy before flattening the composition for distribution. Inspect extracted edges and lettering at delivery size.

Reference documentation: UNI image model · Layers.

HDR, EXR and production delivery

Ray offers HDR generation and sixteen-bit EXR export in ACES2065-1 for colour and compositing work. These are intermediate assets for post-production, distinct from a video ready to play directly on every screen.

Decide the delivery format before generation. A classroom may need a straightforward playback file, while an audiovisual team may require frames with room for lighting and colour adjustments. Check conversion for the destination display: an image that looks correct in the editing environment can appear too dark or washed out in another player. Keep working material and final exports as distinct versions, with identifiable names and settings rather than assuming one format serves every audience.

Reference documentation: Ray3.2 current product · HDR EXR and colour spaces.

Skills and connectors

Skills saves instructions and decisions from a creative process for reuse. Connectors bring in assets from Google Drive, Dropbox, Frame.io, Box and Airtable. Luma describes a read-and-add model that imports and saves work without deleting or overwriting files in connected storage.

A useful automation might prepare three formats of an approved illustration while leaving review pending. Define inputs and details that must remain constant. Test different image proportions and long text: success with one version does not establish that all will work. When sharing a process, include its purpose and acceptance criteria so another person can detect an incorrect output rather than merely confirm that it ran.

Reference documentation: Skills and repeatable workflows · Connectors.

Current plans and legacy licences

Luma App offers Plus, Pro and Ultra for individuals, with Team and Enterprise options. The listed individual plans include commercial use and differ in credit capacity. Older Dream Machine documents describe tiers such as Lite or Unlimited and tie rights to the plan active during generation.

Do not combine the two tables when assessing a purchase or an old asset. A subscription upgrade does not retrospectively convert a personal-use generation into commercial content. Production cost depends on model, resolution, duration and retries; HDR and EXR can also affect consumption. Before commissioning a series, budget for one complete piece including trials, revisions and delivery formats.

Reference documentation: Current plans · Legacy DreamMachine licensing.

API and transition from legacy models

The migration guide moves older Ray integrations to the Luma Agents API with Ray3.2. The requested operation is expressed as generation, editing or reframing. Retirement dates depend on each customer’s notice and should not be inferred from a generic guide.

An integration must retain the job identifier, check status and save results. Migration is more than changing a name: compare duration, references, format and cost. For a material library, use a small representative selection and compare against existing deliveries. Record which versions were recreated and which still use previously approved files. Replacing the model should trigger review of affected outputs rather than silently inheriting previous approval.

Reference documentation: Migration from legacyRayAPI.

Reception and training transparency

In June 2024, The Verge identified a character resembling Mike Wazowski in the fictional Monster Camp trailer and questioned training transparency. This was public criticism rather than a judgment. In October 2026 Luma published a Ray training summary describing licensed, public and synthetic data and an initial third-party weight base.

These documents concern different moments and models. The current summary adds information but does not reconstruct every training example or establish the originality of a specific output. For corporate material, review recognisable characters, brands and similarities as well as permission for supplied references. A historical criticism should neither be ignored nor automatically attributed to every later model.

Reference documentation: Ray training content summary,23October2026 · The Verge historical report,18June2024.

Data and conditions of use

The general terms provide licences to process inputs and outputs, including model development and training. Enterprise commitments must be checked in the relevant contract. Documentation also distinguishes restricted data and requires rights in the material used.

Before uploading workplace images, inspect people, readable documents and details revealing internal information. A depersonalised diagram is often sufficient for a recreation. Do not use health records as production references. Define board access, retention of originals and removal of unnecessary assets. Review should cover the visible file and the text and references accompanying the brief. Account permissions and contractual processing conditions are separate questions that both need a clear owner.

Reference documentation: Terms of Service.

Practical applications in OSH

It can support illustrative scenes for discussions about occupational safety and health. Its potential value is visualising a fictional context and prompting questions about hazards, decisions or communication, while clearly separating creative representation from documentation of actual conditions. The facilitator should connect the discussion back to the real workplace.

It can prepare discussion material, compare visual perspectives or determine framing for a later video. A fictional scene can prompt questions about missing information: visibility, routes, signs, authorisation or site conditions.

It can also draft a narrative later illustrated with authentic footage. Generated clips should not be used to measure safe distances, braking capability or the behaviour of actual equipment. Those questions require information and methods appropriate to the real situation.

Worked example and result checks

Example: prepare a fictional loading area scene to discuss separation between pedestrians and vehicles. A specialist reviews its elements, avoids assigning real measurements to the generated space and uses the scene as a starting point for analysing the actual site.

1. Define the discussion question: what information would be needed to assess pedestrians and vehicles sharing this environment?
2. Prepare a conceptual layout with distinct areas, without inventing measurements or presenting it as a workplace plan.
3. Generate limited movement and review continuity. Check that barriers, routes and positions do not change inadvertently.
4. Identify what cannot be inferred from the image. Prepare questions encouraging participants to request further evidence.
5. Connect the discussion with the actual workplace and approved documentation. Keep the clip clearly identified as an illustrative recreation.

This proposed case teaches questioning and recognition of uncertainty rather than a particular driving manoeuvre.

Limitations and conditions of use

Visual coherence does not guarantee correct distances, speeds or physical behaviour. Do not derive measurements or engineering conclusions from a generated scene. Review continuity and safeguards frame by frame; for an actual manoeuvre, use verified references for the equipment and workplace.

A clip’s beginning and end may appear coherent while intervening motion passes through objects or changes the space. Review the complete sequence at normal speed and pause where an action might be interpreted as instruction.

Extending a video can introduce new details. Do not attribute generated information to the original reference. Keep the original and document which parts were created or transformed, particularly when combining them with authentic recordings.

Reference documentation: Luma — Dream Machine web quick start · Luma — How to use keyframes.

When to choose it and how to complement it

Luma Dream Machine fits visual exploration and illustrative sequences. Google Flow offers another reference-and-scene workflow to compare. VEED can support final editing with recorded footage, wording and captions.

Test a scene containing something that must stay still and another involving a simple transition. Evaluate temporal consistency and correction effort rather than choosing only by the cover image. Define rejection criteria before generation and retain them with the selected clip.

Frequently asked questions

Do keyframes guarantee correct movement?

No. They guide the scene; the generated transition still requires review.

Can it calculate a safe distance?

No. Creative video is neither a measurement nor a validated simulation of equipment or a location.

Can an imperfect scene still be useful?

It may help discuss missing information if clearly identified as illustration and not used to teach an incorrect practice as correct.