How to use Luma Dream Machine: Ray3.2 workflow and prompts

For a first useful test, animate one strong image or describe one short shot. Review motion, identity, object continuity, camera behavior and export quality before building a longer sequence.

Choose the right Dream Machine workflow

• Text-to-video: use it to explore a scene when exact composition is flexible.
• Image-to-video: use it when the subject, product, character or opening composition already exists.
• Keyframes: use start and end frames, or the keyframe controls currently available to the selected model, when the shot needs defined visual beats.
• Modify Video: use an existing clip when its performance or camera motion should guide a transformation.
• Reference mode: use permitted character or visual references when identity or style continuity matters and the selected workflow supports them.

Feature availability, clip length, resolution, credits and controls vary by model, plan and product surface. Check what the composer displays for the selected mode instead of applying one limit to every Dream Machine workflow.

How to create a first video in Dream Machine

1. Create a Board

Open Dream Machine, sign in and create a Board for the project. Luma's web quick-start guide describes Boards as the place where visual projects and their iterations are organized.

2. Define one shot

Write a single visible event with one subject and a clear beginning. Decide the aspect ratio and destination before generation. Split dialogue, location changes and multiple actions into separate shots when they cannot fit cleanly.

3. Choose text, image or source video

Use text for exploration, an image for composition control, keyframes for defined states, or Modify for motion transfer and restyling. Select based on what must stay fixed, not which control sounds most advanced.

4. Select the current model and settings

Record the model name, mode, duration, resolution, aspect ratio, references and any strength control shown in the interface. Luma announced Ray3.2 with multi-keyframe control, longer generations, HDR and EXR capabilities, but a specific option can still depend on the selected access surface and plan.

5. Write the prompt as shot direction

Prompt formula: framing and camera + subject + one action + setting + light and atmosphere + visual treatment + required continuity.

Concrete nouns and observable motion are easier to diagnose than “make it cinematic.” Luma's best-practices guide recommends natural, detailed language and describes camera motion, references, Modify, Extend and Keyframes as separate refinement controls.

6. Generate alternatives without changing everything

Create several candidates with the same brief and settings. Save the candidate, prompt and configuration. If every result fails in the same way, change one cause—source image, action complexity, camera movement or reference strategy—rather than rewriting every variable.

7. Refine with the control that matches the failure

• Composition or identity drifts: start from a stronger image or permitted character reference.
• The ending is uncontrolled: add an end frame or supported intermediate keyframes.
• Motion is useful but the world is wrong: use Modify Video and compare strength settings against the same source.
• The shot ends too soon: use Extend when the source and selected model are eligible, then inspect the seam.
• Only one detail is wrong: modify that element or fix it downstream rather than regenerating an otherwise approved shot.

Luma's current Ray3 Modify guide explains that lower, Adhere-oriented strength preserves more source edges and detail, while higher Reimagine-oriented strength permits broader reinterpretation and can change inferred camera behavior. Its keyframe guide explains start and end frames as visual anchors.

8. Download only after final-file QA

Watch the exported file at normal speed, muted, and frame by frame around failures. Confirm the correct duration, resolution, aspect ratio, color handling and audio behavior. A good preview is not proof that the downloaded deliverable is ready.

When image-to-video is the better starting point

Use an image first when a client-approved product, character, layout or opening frame must survive generation. Prepare the still at the target ratio, remove accidental details, and describe motion instead of repeating everything already visible. Inspect whether the model changes logos, text, hands, product geometry or identity.

How to use keyframes

• Use the start frame to lock the opening composition.
• Use an end frame when the destination matters more than a free continuation.
• Use supported intermediate keyframes to define important beats rather than asking one paragraph to control the full timeline.
• Keep lighting, character and object continuity plausible between anchors.
• Review the path between frames; accurate endpoints do not guarantee coherent motion between them.

Luma's Ray3.2 announcement describes support for up to 16 keyframes and clips up to 20 seconds at 1080p. Treat those as documented model capabilities, then confirm what the current Dream Machine composer or API makes available for the selected account and request.

How to use Modify Video

• Upload a short, high-quality source clip whose motion and framing are worth preserving.
• Add a permitted character or frame reference only when the workflow needs it.
• Describe what should change and what must remain consistent.
• Compare a restrained strength with a broader reinterpretation using the same input.
• Check camera motion, silhouette, background, face, hands, object contact and original audio across the complete output.

Luma's current Ray3 Modify documentation lists a 10-second maximum input for that mode and says it retains original audio. Those limits apply to Modify, not automatically to every Ray3.2 generation path.

Common Dream Machine mistakes

• Calling the app and model the same thing: cite Dream Machine for the product surface and Ray3.2 for the model or API request.
• Writing a whole story as one shot: plan separate clips and define the transition between them.
• Using text-only generation for a fixed product: start from an approved image and verify every brand detail.
• Stacking camera directions: choose one motion that supports the action.
• Changing prompt, model and settings together: keep a comparison log so the cause of improvement is visible.
• Assuming more strength is better: in Modify, broader reinterpretation can sacrifice source geometry or camera preservation.
• Publishing the preview: inspect the final downloaded file, rights, disclosure and destination requirements.

A production-ready review checklist

• The intended action occurs and the camera behaves as requested.
• Faces, hands, bodies, products, text and object contact remain coherent.
• Characters and important objects stay identifiable across frames and shots.
• Keyframe transitions and extensions have no visible jump or unexplained transformation.
• Audio is present or absent as intended and synchronized with the edit.
• The final file uses the required ratio, resolution, duration, color and codec for the destination.
• Every source image, video, person, voice, brand and reference is authorized for the intended use.
• Realistic synthetic or altered media has the context or disclosure required by the destination.

Official sources checked

Workflow and interface guidance comes from Luma's web quick start, best practices, keyframe guide and Ray3 Modify guide. Current model capability language comes from Luma's Ray3.2 announcement. Sources were checked September 13, 2026.