Skip to content
eWiki The practical AI reference

Video Creation

Luma Dream Machine AI Video: Choosing Text, Image, and Generation Controls

Clear glass with blue liquid or light on a pale tabletop, surrounded by stacked white paper, a blank board, and a black ruler.

A concise decision guide for choosing a text prompt or image keyframe and using the officially documented Ray2 generation controls.

Dream Machine and Luma: the naming explained

Luma’s current app page says the experience previously associated with Dream Machine is now part of Luma, where creative agents can take work from a brief through delivery (Luma). The provider’s learning hub still uses the Dream Machine name in its product guides, including a Ray2 video-generation guide published January 28, 2025 (Luma Learning Hub).

That distinction matters when following instructions: the app’s broader name is Luma, while the documented workflow below belongs specifically to the dated Ray2 guide. Treat its labels as a reference point and use the options displayed in your own interface when they differ.

Suggestion: Before starting a project, decide whether your priority is discovering a scene or directing motion from an existing composition. That choice provides a clearer basis for selecting text or an image than beginning with every control at once.

Black camera lens, control panels, pen, blank paper, and a blurred photo arranged on a light tabletop.

Choose text for scene exploration

The official Ray2 guide documents video generation from a written prompt. Its workflow asks the user to describe the video, select Describe, choose a clip length and resolution, select Ray2, and generate the file (Ray2 generation guide).

Suggestion: Choose text when the composition is still open to interpretation. Draft the request around visible information:

  • the main subject;
  • an observable action;
  • the setting and time of day;
  • the intended visual style;
  • one camera movement.

For example, a planning prompt might specify a cyclist crossing a rain-lit street while the camera tracks from the side. This is an editorial example, not a promised output.

Keep the first attempt narrow. If the result suggests the wrong framing, revise the camera or setting before adding more details. Changing one major variable per attempt may make comparisons easier, although it does not guarantee greater consistency.

Person arranging printed images on a white table beside a camera on a small tripod, a rectangular light, and geometric objects.

Choose an image when composition comes first

The provider’s Ray2 guidance says images can be uploaded as starting or ending keyframes, and that keyframes can define the beginning and end of a scene (Luma Learning Hub).

Suggestion: Begin with an image when you already know the subject placement, palette, lighting, or framing you want to preserve as the visual point of departure. Prepare the image at the intended composition before uploading it; then use the prompt to describe movement rather than restating every visible detail.

A useful planning split is:

  • Starting image: identifies where the shot begins.
  • Motion prompt: describes what changes and how the camera moves.
  • Ending image, if used: indicates the destination of the transition.

An ending keyframe adds another constraint. Consider introducing it only when the final composition is important. This is a workflow suggestion, not an assurance that the generated transition will reproduce either frame exactly.

Select generation controls deliberately

The dated Ray2 guide lists 5- or 10-second clip choices and 540p, 720p, or 1080p output. It also documents Extend, Keyframes, and Loop; the guide says extensions can take a five-second clip as far as 30 seconds, with possible quality decline over longer extensions (official Ray2 guide).

Suggestion: Use the shortest clip that can express one action. Treat each generation as a shot rather than a complete sequence, then review subject shape, direction of movement, framing, and the final moment before extending it.

Use a loop only when the action can return naturally to its starting state. Use Extend when continuing the existing motion is more important than designing a separate shot. Choose resolution according to the evaluation stage: a lower setting may suit early comparisons, while a higher setting may suit a selected version. Those are production suggestions, not claims about speed, cost, or output quality.

A practical text-versus-image workflow

Suggested workflow: First, write a one-sentence shot objective. Next, decide whether a written description contains enough visual direction. If not, prepare a starting image. Generate one short attempt, inspect only the largest mismatch, and revise that variable. Add an ending keyframe, looping, or extension only after the basic shot direction is acceptable.

For text-led work, prioritize the subject, action, setting, and camera cue. For image-led work, let the still carry composition and use the prompt chiefly for movement. In both cases, save the prompt and note which input changed between attempts. This record can make subjective comparisons more disciplined without implying that identical settings will always produce the same result.

Avoid planning a long sequence around a single generation. A shot-by-shot outline gives each clip one purpose and makes it easier to decide whether the next attempt needs a revised prompt, a stronger reference image, or a separate shot.

Conclusion

For Luma Dream Machine AI video, my practical recommendation is to start with text when you are exploring the scene and to start with an image keyframe when composition is already decided. Keep the first generation short, specify one clear action and camera direction, and evaluate it before adding an ending keyframe, Loop, or Extend.

The official app now presents Dream Machine’s broader experience under Luma, while the detailed controls cited here come from Luma’s January 2025 Ray2 guide (Luma app; Ray2 guide). Use those documented options as a planning framework and follow the controls shown in the version available to you.

Sources

Disclosure

This article was prepared with AI assistance.

Improve this guide

Report a problem or suggest the next question

Submissions enter a private editorial queue. They are reviewed and deduplicated before any public update or new page is created.