Video

Luma Ray 3: how to write prompts the model actually understands

Luma · Updated:

Luma Ray 3 is the Ray 3 lineup: Ray 3.14 (the workhorse, default for 90% of tasks) and Ray 3 Reasoning (the multimodal «thinking» model). Available in Dream Machine and Adobe Firefly. Unique features: Character Reference for character identity, Draft Mode (5–10x faster), 16-bit HDR, Visual Annotation, native 1080p. Duration is 5 or 10 seconds.

What's new in Ray 3

Ray 3 is split into two models. Ray 3.14 is the workhorse — the fastest in the lineup, recommended as default for 90% of tasks. Native 1080p, HDR, EXR export. Excellent cinematic quality: filmic lighting, stable camera, convincing motion.

Ray 3 (Reasoning) is the first multimodal «thinking» video model. It breaks down complex creative briefs into steps, like a director planning a storyboard. It can evaluate and refine its own results for logical consistency.

Unique features in Ray 3 only (not in 3.14): Character Reference — upload a reference image to preserve identity across T2V, I2V, V2V and Reference Mode. Draft Mode — generation 5–10x faster and cheaper for idea iteration. 16-bit HDR — native extended dynamic range for studio quality. Visual Annotation — draw instructions directly on the start frame for precise motion control.

  • Ray 3.14 — default for 90% of tasks, native 1080p
  • Ray 3 Reasoning — Character Reference, Draft Mode, 16-bit HDR
  • Visual Annotation — instructions drawn on the frame
  • EXR export for professional color grading (540p/720p)
  • Duration 5 or 10 seconds, Modify up to 18s, Extend up to ~30s

Prompt structure

The Ray 3 formula is identical to the rest of the family: [Subject] + [Mid-Action Verb] + [Setting/Environment] + [Secondary Motion/Consequences] + [Camera Movement] + [Lighting/Mood].

Like Ray 2, Ray 3 is a «positive only» model. Negative prompts are counter-productive — describe only what you want. Use present continuous: «running», «pouring», «spinning» — not «begins to», «will», «starts». Ray 3 defaults toward a cinematic look — if you want documentary or social-style, state it explicitly.

Optimal length is around 100 words focused on the action. Under 15 words — the model fills in too much. Over 200 words — overload. Lighting and mood significantly drive quality: «golden hour side lighting», «moody noir lighting, high contrast», «crisp commercial lighting».

Character Reference (Ray 3 only)

The headline Ray 3 feature is Character Reference for preserving character identity across multiple generations. Upload a reference image and the prompt describes only the SCENE and ACTION — appearance is locked by the reference.

Key rule: do NOT re-describe the character's appearance. The reference image does that. The prompt «Character walking through a foggy forest, leaves crunching underfoot, mist swirling around legs, low-angle tracking shot» is correct. The prompt «Tall woman with blonde hair, blue eyes, wearing red jacket, walking through a foggy forest» re-describes what the reference already defines and throws the model off.

Best results come from clear, well-lit photos showing the face and key features. Character Reference works in T2V, I2V, V2V and Reference Mode. For multi-shot narratives with one character — must-have. For one-off shots without identity needs — not required.

Draft Mode and generation modes

Draft Mode (Ray 3 only, not in 3.14) — fast low-resolution generation for idea iteration. 5–10x faster and cheaper. Strategy: test ideas in Draft, switch to standard 1080p mode for final renders. Especially useful for prompt debugging and finding the right composition.

Modify (V2V) — transforming existing video, up to 18 seconds. Three modes: Adhere (1–3, preserves the original), Flex (1–3, balance), Reimagine (1–3, creative freedom). Describe the END STATE, not commands: «Cyberpunk neon city at night, rain-slicked streets, purple and blue lighting» works; «Change the sky to blue» doesn't. Start at Adhere 1–2 and raise as needed.

Extend pushes total length up to ~30 seconds, Loop produces a seamless looped video. Visual Annotation — drawing instructions directly on the start frame: motion zones, object positions, camera paths. A visual supplement to the text prompt.

Common mistakes

  1. 1. Re-describing appearance with Character Reference

    When using Character Reference, don't describe the character's appearance in the prompt — the reference image does that. The prompt should focus on the SCENE and ACTION. Describing appearance duplicates the reference and throws the model off — it tries to combine two sources of character information, producing instability.

  2. 2. Forbidden words in Ray 3

    «Vibrant», «whimsical», «hyper-realistic», «beautiful», «amazing», «stunning» empirically degrade quality across all Ray models, including Ray 3. Replace with concrete description: instead of «hyper-realistic» write «sharp focused detail, cinematic film grain». Instead of «beautiful» use concrete light, composition, mood description.

  3. 3. Commands and temporal language in Modify

    «Change the sky to blue», «transforms into a forest», «remove the clouds» — don't work in Modify. The fix: describe the END STATE as a standalone scene. «Clear blue sky with bright daylight» works; «change to blue sky» doesn't. No temporal language («changes to», «transforms into») — the model doesn't understand sequence.

  4. 4. Prompt too short or too long

    Under 15 words — the model fills in too much, results are unpredictable. Over 200 words — detail overload, the model loses focus. ~100 words is the gold standard for Ray 3: enough detail on subject, action, scene, secondary motion, camera, and light without overload.

  5. 5. Skipping Draft Mode for iteration

    A full Ray 3 render at 1080p costs time and credits. Draft Mode (5–10x faster) is purpose-built for idea iteration and prompt debugging. Strategy: cycle variants in Draft until stable, switch to standard mode only for finals. Don't iterate at full quality — that's technical debt.

Before / after examples

Example 1

Before

young woman hiking, will look beautiful, hyper-realistic

After

Young woman in red jacket hiking up a mountain trail, backpack bouncing slightly with each step, hair tied back, breathing visible in cold air, camera tracking alongside, golden hour side lighting, crisp cinematic detail.

Forbidden words («will look beautiful», «hyper-realistic») removed. Added secondary motion (backpack bouncing, breathing visible), present continuous (hiking), explicit camera (tracking alongside) and light (golden hour side lighting).

Example 2

Before

Tall woman with blonde hair walking through forest (with Character Reference)

After

Character walking through a foggy forest, leaves crunching underfoot, mist swirling around legs, low-angle tracking shot, soft diffused morning light, atmospheric depth.

With Character Reference do NOT re-describe appearance — that's what the reference image handles. The prompt focuses on the scene (foggy forest, mist), action (walking, leaves crunching), and camera (low-angle tracking). This gives the best consistency across shots.

Example 3

Before

Change the cloudy sky to clear blue (Modify)

After

Clear blue sky with bright daylight, soft white clouds at the horizon, warm sunlight casting natural shadows — same composition and subject as original, Adhere intensity 2.

For Modify in Ray 3, describe the END STATE as a standalone scene, not as a command. Specifying intensity (Adhere 2) and anchoring «same composition» yields predictable results with minimal drift.

Frequently asked

What's the difference between Ray 3 and Ray 3.14?
Ray 3.14 is the workhorse — the fastest in the lineup, recommended for 90% of tasks: fast generation, native 1080p, HDR, EXR export. Ray 3 (Reasoning) is the multimodal «thinking» model with unique features: Character Reference for character consistency, Draft Mode (5–10x faster), 16-bit HDR, Visual Annotation. Use Ray 3.14 by default, switch to Ray 3 when you need identity across shots or fast iteration.
What is Character Reference in Ray 3?
Character Reference is a Ray 3-only feature (not in 3.14) that lets you upload a reference image of a character to preserve identity across multiple generations. Works in T2V, I2V, V2V and Reference Mode. Key rule: the prompt describes only the SCENE and ACTION — appearance is set by the reference. Do NOT re-describe appearance. Best results with clear, well-lit photos.
When should I use Draft Mode?
Draft Mode (Ray 3 only) is for idea iteration and prompt debugging. 5–10x faster and cheaper than standard rendering. Strategy: test the prompt in Draft until stable, switch to standard 1080p mode for finals. Don't use Draft for final renders — quality is lower. Don't use standard mode for every iteration — that's technical debt.
What is Visual Annotation?
Visual Annotation is a Ray 3 feature that lets you draw instructions directly on the start frame: motion zones, object positions, camera paths. It's a visual supplement to the text prompt, especially useful for complex motion, precise object placement, and multi-subject scenes. It doesn't replace text — it complements it.
Does Ray 3 support 16-bit HDR?
Yes, Ray 3 is Luma's first model with native 16-bit HDR. This is extended dynamic range for studio quality: dramatic lighting (sunsets, neon, fire, stage lights) looks convincing. EXR export (540p and 720p) gives uncompressed HDR frames for professional color grading in DaVinci Resolve, Nuke, and other post tools. Ray 3.14 also supports HDR.
How long can I generate?
T2V and I2V — 5 or 10 seconds per run. Modify (V2V) — up to 18 seconds. Extend pushes total length to ~30 seconds, beyond that quality degrades. Each Extend run is a separate prompt with new content. For seamless looping use Loop. For narratives longer than 30 seconds — multiple separate shots tied together via Character Reference.
Does Opten support Luma Ray 3?
Yes, the Opten extension auto-detects every model in the Ray 3 lineup (Ray 3, Ray 3.14, Ray 3 Reasoning) and scores prompts with their specifics in mind. It checks correct use of Character Reference (no re-describing appearance), absence of commands in Modify, present continuous, concrete camera and light descriptions, and the optimal ~100-word length. One click gives you a rewrite targeted at the selected model.

Related models

Ready to write Luma Ray 3 / Ray 3.14 / Ray 3 Reasoning prompts in one click?

  • Auto-detects the model inside its native interface
  • Scores every line of your prompt
  • One-click rewrite into the correct structure
ChromeYandex BrowserChrome / Yandex BrowserInstall extension

Pro — $2.99/month or ₽199/month · cancel anytime

Stop Guessing. Generate
On The First Try.

Install Opten in 30 seconds and score your next prompt.

Opten is a Chrome extension that scores AI prompts for the specific model. Supports 60+ image and video models — Midjourney, GPT Image 2, Kling, Sora, Nano Banana, Flux — and rewrites them in one click inside the Syntx, Higgsfield, and Freepik interfaces. From $2.99/month.

© 2026 Opten · IE Nikolai Shupletsov · Tax ID 306389672