Video · guidance updated 22 Nov 2026
Luma (Dream Machine / Ray) prompt guide
Ray 3.2: V2V production workhorse, 20s max source duration. Multi-keyframe i2v up to 64 keyframes at arbitrary frame indexes. Supports HDR, EXR export, Adherence (Motion/Structure 1–9), Character controls (Faces/Bodies/Poses). Start_frame/end_frame limited to 5s, no HDR.
How to prompt Luma (Dream Machine / Ray)
- Confirm goal, deliverable, usage, and priorities with agent before generating—have it summarize understanding and ask clarifying questions.
- Choose input strategy based on available assets: pure T2V, I2V with start/end anchors (5s only), V2V from source video, or multi-keyframe i2v for strongest control across arbitrary moments.
- For V2V, describe only target end-state look; Ray 3.2 preserves source motion and timing exactly, output duration matches source (up to 20s).
- Pin keyframes at exact source frame indexes before motion generation to lock style and consistency; use 1+ guide frames for each distinct moment.
- Create keyframes by exporting source frames, art-directing in Photoshop/Nuke or Luma's image tools, then feeding back at same indexes for production restyles.
- Use Adherence controls (Motion 1–9, Structure 1–9) to protect critical elements; increase Motion if movement drifts, increase Structure if layout changes too much.
- Toggle Character controls (Faces, Bodies, Poses) only for attributes that must survive; disable what you want freed to transform.
- Break longer content into 4–6 discrete moments; generate each with dedicated keyframes and focused prompt for consistency.
- Write prompts as declarative end-state descriptions ('brushed steel', not 'change to metal'); specify concrete camera, lighting, materials, and palette.
- Use positive phrasing only; remove negation ('no', 'not', 'without') from all prompts.
- Diagnose issues when iterating—classify into motion, camera, style drift, character drift, pacing—and fix highest-impact problems first.
- Generate multiple options in batches, compare side-by-side, and provide focused feedback citing what works, what doesn't, and specific adjustments needed.
- Export, art-direct, and iterate keyframes using external tools, then re-feed at same source indexes; use Enhance and inference modes (Speed vs Quality) to refine results.
The shape of a good prompt
V2V: [Source video (up to 20s) + keyframes at frame indexes] → prompt [target look: materials/palette/light/style]. Adherence: [Motion 1–9] [Structure 1–9]. Characters: [toggle Faces/Bodies/Poses on/off]. Multi-keyframe i2v: [keyframes[] + keyframe_indexes[]] + prompt [camera] of [subject] [action] in [environment]. I2V (5s only): [start_frame/end_frame] + prompt. T2V: [prompt describing camera, subject, materials, action, environment, light, palette].
Avoid
- Imperative edit commands ('make', 'turn', 'change', 'fix') in V2V prompts.
- Describing sequence of changes or story progression in V2V; prompt must express end-state only.
- Negation ('no', 'not', 'without', 'avoid') anywhere in prompts.
- Generic adjectives ('cinematic', 'dreamy', 'high quality', 'perfect') instead of concrete visual specifications.
- Generating full 60–90s or long-form content in one pass; break into 4–6 moments with individual keyframes.
- Skipping keyframes when consistency across moments matters; always pin guide frames at critical beats.
- Using start_frame/end_frame (legacy path, 5s max, no HDR) when multi-keyframe i2v or V2V would provide stronger control.
- Treating agent as vending machine; always define goal, confirm understanding, provide structured feedback, and iterate diagnostically.
Example
Request a runner in Tokyo at night
Slow tracking shot alongside a female runner in a black windbreaker on a quiet Tokyo backstreet at night. Wet asphalt mirrors soft pink and cyan neon signs; light rain falls through the glow of a single streetlamp. Shallow depth of field, anamorphic lens flares, teal-and-magenta palette, cinematic 35mm film grain.