Luma Ray 3.2 Prompt Templates (30 Copy-Paste Prompts, 2026)

TL;DR: Luma Ray 3.2 prompts work best in six ordered slots: subject, action, camera move, lens and framing, light, and finish. Ray 3.2 runs four workflows from one model, and each needs a different prompt shape. Below are 30 copy-paste templates, plus the six places Luma's own published specs disagree with each other.

What are the best Luma Ray 3.2 prompt templates?

The best Ray 3.2 prompt is one sentence of subject and action, followed by one clause each for camera move, framing, light, and finish. That order matters. Put the camera move early so motion gets resolved before styling.

Luma's own documented examples follow exactly that shape. Their text-to-video sample in the Agents API guide is "A slow dolly shot through a misty greenhouse at sunrise", which is camera move, subject, environment, light, in twelve words. Their image-to-video sample is "The character turns to face the camera and smiles". Neither is long. Neither uses keyword stuffing.

Here is the skeleton every template on this page fills in:

[SUBJECT doing ACTION], [CAMERA MOVE], [SHOT SIZE + LENS],
[LIGHT + TIME OF DAY], [FINISH / FILM STOCK / GRADE]

That is a prompt template in the strict sense: fixed slots, swappable contents. It is deliberately not JSON. Ray 3.2's prompt field is plain prose of 1 to 6,000 characters, and Luma's prompting guide tells you to write like a creative director briefing an artist rather than to stuff keywords.

If you want the reasoning behind each slot, the mechanics of frame grids, and what to do when a generation ignores you, that is the companion piece: how to prompt Luma Ray 3.2. This page is the template pack. Take the blocks and go.

Which Ray 3.2 workflow should each prompt target?

Ray 3.2 is four workflows in one model, and a prompt written for one will underperform on another. Text-to-video invents everything. Image-to-video interpolates between images you supply. Modify Video re-renders footage you already shot. Reframe outpaints a clip into a new aspect ratio.

Two facts change how you write for all four. First, there is no native audio on text-to-video or image-to-video, per Luma's Ray 3.2 FAQ, so describing sound design burns characters for nothing. Modify Video and Reframe preserve your original audio instead. Second, there is no negative prompt. The Agents API FAQ is explicit that no negative_prompt parameter exists and that you should describe what you want positively.

Where do Luma's own published numbers disagree?

This is worth knowing before you trust any spec you read about Ray 3.2, including ours. On August 26, 2026 we checked Luma's marketing pages against Luma's developer docs and found six live contradictions. Where they persist, we give both figures with their sources rather than picking a winner.

None of these are catastrophic, and a couple have plausible reconciliations. The 20s Modify figure is an output length at 24fps while the 18s figure is an input cap, so a source can be shorter than what comes out. The keyframe gap most likely reflects a UI limit sitting inside a wider API ceiling. But Luma does not say so anywhere, and two of the reframe figures sit in different answers on the same FAQ page.

The practical rule: build to the smaller number. Sixteen keyframes, a 10-second reframe source, a 720p target. Then test upward.

It is worth being clear about why this matters for a template pack. A template is a promise that the block will run as written. If we wrote "pin sixteen keyframes across a twelve-second reframe" on the strength of the marketing FAQ, roughly half the blocks below would return a 400 against the current API. Every limit quoted from here on is the developer-docs figure, because that is the one the endpoint enforces.

Product and commercial templates

Six blocks for product, packaging and e-commerce work. Ray 3.2 accepts six aspect ratios for video: 9:16, 3:4, 1:1, 4:3, 16:9 and 21:9. Pick the delivery ratio at generation time rather than reframing later, because reframe costs a second pass.

1. Hero orbit on a clean cyc

A [product] standing upright on a matte concrete plinth, a slow 180-degree
orbit around the product from left to right, medium close-up on a 50mm lens,
soft top key with a large white bounce card camera-left and a deep falloff
background, clean commercial finish with a shallow depth of field

2. Slow push-in reveal

A [product] centred on a dark reflective surface, a slow push-in that ends
tight on the [label / seam / logo], macro framing on an 85mm lens,
single hard key raking across from the right to catch the edge highlight,
high-contrast product finish with crisp specular detail

3. Material and texture macro

Extreme close-up of [material] on a [product], a slow lateral drift across
the surface, macro lens at f2.8 with the focus plane held on the texture,
low-angle grazing light that exaggerates the grain, natural colour, no grade

4. In-hand demonstration

A pair of hands lifting a [product] out of its box and turning it once toward
camera, locked-off camera with the hands entering frame from the bottom,
medium shot on a 35mm lens, soft overhead diffusion and a warm practical
behind, clean lifestyle finish

5. Product in use, lifestyle

A [person] using a [product] on a [surface] in a [room], a gentle handheld
drift that follows the hands, medium wide on a 28mm lens, late-afternoon
window light from camera-left with visible dust in the air,
warm documentary finish on 35mm film stock

6. Liquid and pour

A stream of [liquid] pouring into a [vessel] and settling, a slow rise from
below the rim to eye level, tight close-up on a 100mm macro, backlit through
a diffusion panel so the liquid reads translucent, high-speed feel with
crisp motion and a neutral grade

Cinematic and camera-move templates

Seven blocks built around a single named camera move each. Ray 3.2 responds better to one clearly stated move than to three stacked on top of one another. If you want a compound move, put the beats on keyframes instead, which is group four below.

7. Slow dolly through an interior

A slow dolly forward through [an interior space] at [time of day],
the camera passing [a foreground object] on the left,
wide shot on a 24mm lens at eye height, low ambient light with a single
practical deep in frame, moody cinematic finish with lifted blacks

8. Tracking shot following a subject

[Subject] walking away from camera along [a path], the camera tracking
behind at a constant distance and matched pace, medium shot on a 40mm lens
at chest height, overcast flat daylight, naturalistic finish with muted colour

9. Crane reveal

The camera rises from ground level to a high vantage over [a location],
revealing [what appears at the top of the move], starting close and
ending wide on a 35mm lens, golden-hour backlight with long shadows,
epic cinematic finish with a warm highlight roll-off

10. Handheld documentary

[Subject] doing [action] in [environment], loose handheld camera with small
natural corrections and a breathing frame, medium close-up on a 50mm lens,
available light only from [source], gritty 16mm documentary finish with
visible grain

11. Locked-off frame, motion inside it

A completely static locked-off camera on [a composition],
[what moves] moving through the frame from [direction] to [direction],
wide shot on a 28mm lens, unchanged light throughout the take,
clean observational finish

12. Aerial push

An aerial view descending toward [a location], the camera pushing forward
and down at a steady rate, very wide on a 20mm lens, early-morning haze
with the sun low on the horizon, high-altitude cinematic finish with
atmospheric depth

13. Rack focus between planes

[Foreground subject] in soft focus with [background subject] sharp behind,
the focus racking from back to front over the length of the shot,
medium close-up on an 85mm lens at f1.8, single soft key from camera-right,
shallow anamorphic finish with oval bokeh

The camera vocabulary that Ray 3.2 reliably reads is small and worth memorising: dolly, track, orbit, push, pull, rise, descend, pan, tilt, handheld, locked-off, rack focus. Compound and invented moves degrade fast. The same discipline applies across models, which is why our cinematic camera prompts for Veo 3 and Kling uses a near-identical vocabulary list.

Motion-transfer and Modify Video templates

Modify Video re-renders footage you already have while preserving its motion and timing. Luma's FAQ says a restyle keeps the original performance, motion and timing, and preserves lip sync when face tracking is on. So your prompt should describe only the new look, never the motion. The motion is already in the source.

14. Restyle, keep the performance

Transform the scene into [look: e.g. moonlit 35mm film footage], keeping the
same actor, the same blocking and the same timing. Change only the lighting,
the colour and the surface treatment.

15. Environment change

Replace the environment with [new setting], keeping the subject, their
position in frame and their movement exactly as shot.
The new environment is lit to match the existing key direction.

16. Relight to a new time of day

Relight the scene as [time of day / condition], with the key coming from
[direction] and matching ambient fill. Preserve every surface, prop and
camera move; change only the light and the resulting colour temperature.

17. Character transformation

Transform the person into [new character], preserving their facial expression,
their gesture and their posture frame for frame. Keep the performance;
change the identity, the costume and the material treatment.

18. Wardrobe or product swap

Replace [the existing item] with [the new item] on the same subject,
matching the original contact points, occlusion and shadow.
Everything else in the frame is unchanged.

19. Camera-motion transfer

Keep the camera move and its timing exactly as in the source,
and rebuild the world it moves through as [new setting and style].
Subject blocking follows the original.

20. Add a visual effect

Add [effect: e.g. drifting embers, volumetric haze, falling snow] to the scene,
integrated with the existing light direction and the existing camera move.
Leave the subject, the performance and the grade untouched.

Two API-side notes for Modify Video. Start with auto_controls: true, which Luma's editing guide calls the recommended default, and only reach for manual conditioning when auto mode is not enough. When you do, the strength enum runs adhere_1 through adhere_3 for tight preservation, flex_1 through flex_3 for balanced, and reimagine_1 through reimagine_3 when the prompt should drive. Luma's docs name flex_2 as the common default.

{
  "model": "ray-3.2",
  "type": "video_edit",
  "prompt": "Transform the scene into moonlit 35mm film footage",
  "source": { "generation_id": "<prior-generation-id>" },
  "video": {
    "resolution": "720p",
    "edit": { "strength": "flex_2" }
  }
}

Keyframe-driven templates

This is the workflow Ray 3.2 was built for, and it is where prompts stop being the main lever. You pin guide images at chosen frame positions and the model interpolates between them. Frame indexes sit on a duration × 24fps grid: a 5-second clip runs 0 to 120, a 10-second clip runs 0 to 240.

The number of keyframes you actually want is lower than the ceiling. Three across five seconds gives the model roughly two seconds of runway between anchors, which is enough for a camera move to breathe. Push to eight on a five-second clip and you are effectively cutting, not generating: the interpolation windows get short enough that the result reads as a slideshow with motion blur. Start at three, add a fourth only where a beat is landing wrong, and keep the prompt describing the through-line rather than narrating each frame you already pinned.

21. Three-beat story, 5 seconds

{
  "model": "ray-3.2",
  "type": "video",
  "prompt": "A hot-air balloon drifts across the valley as the sun rises",
  "aspect_ratio": "16:9",
  "video": {
    "resolution": "720p",
    "duration": "5s",
    "keyframes": [
      { "url": "https://example.com/beat-1-launch.jpg" },
      { "url": "https://example.com/beat-2-midflight.jpg" },
      { "url": "https://example.com/beat-3-sunrise.jpg" }
    ],
    "keyframe_indexes": [0, 60, 120]
  }
}

22. Six-beat story, 10 seconds

{
  "model": "ray-3.2",
  "type": "video",
  "prompt": "[One sentence describing the through-line, not the beats]",
  "aspect_ratio": "16:9",
  "video": {
    "resolution": "720p",
    "duration": "10s",
    "keyframes": [
      { "url": ".../01.jpg" }, { "url": ".../02.jpg" },
      { "url": ".../03.jpg" }, { "url": ".../04.jpg" },
      { "url": ".../05.jpg" }, { "url": ".../06.jpg" }
    ],
    "keyframe_indexes": [0, 48, 96, 144, 192, 240]
  }
}

23. Look shift at the midpoint

Pin the same composition twice, graded differently, and let the model find the transition:

[Subject and setting], holding the composition steady while the light
changes from [look A] to [look B] across the shot.
No camera move, no blocking change.

Anchor at indexes [0, 60, 120] with look A, a blend frame, and look B.

24. End-card resolve

[Subject and action], settling to a clean hold on the final frame
with [the logo / product / title] centred and static.

Anchor only the last frame at index 120 for a 5-second clip. Everything before it is generated.

25. Keyframed edit on existing footage

{
  "model": "ray-3.2",
  "type": "video_edit",
  "prompt": "[The new look, described once]",
  "source": { "generation_id": "<prior-generation-id>" },
  "video": {
    "resolution": "720p",
    "edit": {
      "keyframes": [
        { "url": ".../open.jpg" },
        { "url": ".../mid.jpg" },
        { "url": ".../close.jpg" }
      ],
      "keyframe_indexes": [0, 45, 90]
    }
  }
}

Note the difference: on video_edit, indexes are positions in the source video's frame grid, not the output grid. Luma's editing guide calls keyframes "the single biggest lever on output quality" for anything past a quick restyle.

Extend and reframe templates

26. Forward extend

Continue the camera move into [what comes next], holding the same lens,
the same light and the same pace.

Pass the finished clip's generation id as video.start_frame.generation_id on a fresh type: "video" request. The prior clip becomes the start and the new generation continues after it.

27. Backward extend

The moments leading up to this shot: [what precedes it],
resolving into the existing opening frame with matched light and lens.

Same request, but pass the id as video.end_frame.generation_id. The new generation is prepended.

28. Perfect loop

[Subject] performing [a cyclical action] against [a static background],
locked-off camera, unchanging light, the end state identical to the start state.

Set video.loop: true. It is create-only, and Luma's docs reject it alongside duration: "10s", hdr: true, end_frame and keyframes.

29. Landscape to vertical reframe

Extend the scene naturally above and below the existing frame,
continuing [the environment] with matched light, texture and perspective.
Add no new subjects.

30. Reframe with explicit source placement

{
  "model": "ray-3.2",
  "type": "video_reframe",
  "prompt": "Extend the scene into cinematic widescreen",
  "aspect_ratio": "21:9",
  "source": { "generation_id": "<prior-generation-id>" },
  "video": {
    "resolution": "720p",
    "source_position": {
      "x_norm": 0.25, "y_norm": 0.0,
      "w_norm": 0.5, "h_norm": 1.0
    }
  }
}

source_position is a normalised rectangle in the output canvas. All four fields are required when it is present, and it is rejected on any type other than video_reframe. Omit it and Luma centre-fits the source for you.

How do you adapt these templates without breaking them?

Swap the bracketed contents, keep the slot order, and change one variable per test. Ray 3.2 is not deterministic, so if you change three things at once you will not know which one moved the output.

Four adaptation rules that hold across all thirty blocks:

1. One camera move per prompt. Stack beats on keyframes, not in the sentence.
2. Describe the light by direction and quality, not by a mood word. "Hard key from camera-right" beats "dramatic lighting".
3. Name the finish explicitly. Luma's prompting guide says to name styles rather than hoping the model infers them. "35mm film stock, lifted blacks" is a finish. "Cinematic" is not.
4. Delete anything the workflow already controls. On Modify Video, motion is preserved by the source, so describing it is wasted. On reframe, the prompt only fills the newly exposed canvas.

The failure mode we see most often is length. Luma allows 6,000 characters and their own FAQ notes that most production prompts sit comfortably under 1,000. Every block above is well inside that. When a shot is not landing, the instinct is to add another clause, and that is almost always the wrong move: the extra clause competes with the camera direction rather than reinforcing it. Cut back to the six slots, fix the one that is wrong, and regenerate.

Then keep the winners. A template you reuse fifty times is worth a slot in a real library rather than a note file, which is the case we make in how to build a personal AI prompt library. The same discipline that applies to text prompts applies here, and the pattern of parameterised blocks with swappable slots is the one running through our 100+ ChatGPT prompt templates too.

Prompt Architects sits on the prompt side of this. We do not generate video and never will. What we do is turn a rough idea into a structured brief with the slots filled, store the version that worked, and let you swap the bracketed variables per client or per campaign with structured output you can paste straight into Luma. The rendering is Luma's job.

What do these templates cost to run?

Ray 3.2 is priced two ways depending on where you run it, and the numbers below were read from Luma's own pages on August 26, 2026. Confirm before you budget, because Luma's API pricing page states that video rates are subject to change ahead of general availability.

HDR doubles the standard rate and HDR plus EXR triples it, on both surfaces. The app plans start at Plus, $30 per month for 10,000 credits, with Pro at $90 and Ultra at $300.

Two things follow from that table. Modify Video costs roughly three and a half times a fresh generation at the same resolution and length, so restyling is not the cheap option. And 360p exists as a draft tier at a fifth of the 720p rate on the API, which makes it the right place to iterate on a prompt before you commit. Just remember that 360p is one of the six specs Luma's own pages disagree about: the models page and the FAQ resolution answer both list only 540p, 720p and 1080p.