Luma Ray: how to write prompts the model actually understands

What the Ray family does

All Luma Ray models support text-to-video, image-to-video with keyframes (start + end frame), extend for continuation, modify for transforming existing video, and loop for seamless looping. Aspect ratios: 9:16, 3:4, 1:1, 4:3, 16:9, 21:9.

Key feature of the family: Ray models are trained directly on video data, so they understand natural motion and physically correct interactions. They do NOT accept negative instructions — describe only what you want to see. The words «vibrant», «whimsical», «hyper-realistic», «beautiful», «amazing», «stunning» degrade quality — empirically confirmed by Luma.

• 5–10 seconds per run, extend up to ~30 seconds
• 720p and 1080p native, 4K via upscale
• Keyframes: start frame + end frame for I2V
• Loop for seamless product videos
• Positive-only model — no negative prompts

Prompt structure

Universal formula across all Ray models: [Subject] + [Mid-Action Verb] + [Setting/Environment] + [Secondary Motion/Consequences] + [Camera Movement] + [Lighting/Mood].

Key rule — present continuous. Use mid-action verbs: «running» instead of «begins to run», «pouring» instead of «starts to pour», «spinning» instead of «will spin». The model doesn't understand future tense or action sequences.

Optimal prompt length is around 100 words focused on the action. Under 15 words — the model fills in too much and results are unpredictable. Over 200 words — detail overload and the model loses focus.

Camera and secondary motion

A concrete camera description significantly improves the result. Movement categories: dolly forward/back, tracking shot, orbit, crane shot, push in, pull out, pan left/right, tilt up/down. Shot size: close-up, extreme close-up, medium shot, wide shot, establishing shot, macro. Angle: low angle, high angle, bird's eye view, Dutch angle, POV, over-the-shoulder.

Secondary motion is the consequence of the main action that makes video feel alive: «ears flapping in the wind», «dust particles catching golden hour sunlight», «fabric billowing outward», «hair flowing», «reflections shimmering on wet pavement», «ripples spreading on water». Without them, video looks static even with a moving subject.

Avoid conflicting camera moves (zoom + pan + orbit simultaneously) — the model tries to execute everything at once and produces chaos.

Modes and keyframes

Text-to-Video — generation from text, the main mode. Image-to-Video — animating an image or transitioning between start/end keyframes. With keyframes the aspect ratio is taken automatically from the uploaded image. Describe only what CHANGES — don't re-describe static elements.

Extend — continuation of existing video. Limit ~30 seconds total, beyond that quality drops. Each extension is a separate prompt describing the new content.

Modify (V2V) — transforming existing video by prompt. Three modes: Adhere (preserve original, minimal change), Flex (balance), Reimagine (freedom). Critical rule: describe the END STATE, not commands. «Cyberpunk neon city at night, rain-slicked streets» works; «change the sky to blue» doesn't.

Loop — seamless looped video for product showcases. Activated via the ∞ icon in the prompt box.

Common mistakes

1. 1. Forbidden words in the prompt

   «Vibrant», «whimsical», «hyper-realistic», «beautiful», «amazing», «stunning» empirically degrade Ray quality. These words give the model no visual information, take up prompt space, and reduce results. Replace with concrete description: «warm orange light», «soft pastel palette», «sharp detail», «cinematic lighting».

2. 2. Temporal phrases and future tense

   «Begins to run», «starts to spin», «will pour», «then transforms» — Ray doesn't understand temporal sequence and works in present continuous. Use mid-action verbs: «running», «spinning», «pouring», «transforming». For sequential actions, split into separate prompts with extend.

3. 3. Negative prompts

   «No text», «without people», «remove watermark», «no clouds» — Ray is a «positive only» model. Negative instructions are either ignored or inverted. Replace with positive description of the desired result: instead of «no clouds» write «clear blue sky»; instead of «without people» write «empty street».

4. 4. Prompt too short

   A prompt under 15 words gives the model too little information and it makes up most of the scene. Optimal length is around 100 words with subject, action, setting, secondary motion, camera, and light. Over 200 words leads to detail overload.

5. 5. Re-describing static elements with keyframes

   When using image-to-video with keyframes, don't re-describe what's already in the frames. Describe only the CHANGES between start and end frame: «hair starts to flow», «camera dollies forward», «light shifts from cool to warm». Detailed descriptions of static elements confuse the model.

Before / after examples

a beautiful golden retriever runs through a field, stunning visuals

A golden retriever running through a wheat field at sunset, ears flapping in the wind, dust particles catching golden hour sunlight, camera tracking alongside at the dog's level, warm cinematic light.

espresso pouring, no spills, no mess

Espresso pouring into a white ceramic cup, dark liquid swirling and forming crema, steam rising slowly, macro close-up, soft warm morning light from the left, shallow depth of field.

change the sky from cloudy to blue (for Modify)

Clear blue sky, bright daylight, soft white clouds at the horizon, warm sunlight, golden hour atmosphere — same composition as original.