Luma AI Text to Video: Prompts, Models and Credit Costs

Text to video in Luma AI starts with a written prompt and ends with a 5- or 10-second clip from Ray3.2, Luma's own video model, released in June 2026. On Luma's pricing table a 5-second Ray3.2 clip at 1080p costs 400 credits in SDR, a 5-second 360p draft costs 20, and HDR doubles the base price.^[1] The catch is sound. Ray3.2 clips are silent, so you either add audio inside the workspace or write the prompt for a model that renders sound in the same pass. The verdict: Luma suits people who want to draft cheaply, choose a model per shot and finish in 1080p HDR. If you want one button that returns a narrated video, plan for more steps.

Step 1: Choose the model before you write a word

The prompt you write depends on the model that reads it, so this choice comes first. Luma's text-to-video page says Ray works out composition, motion and scene logic before it renders a frame, with native 1080p and HDR output; Luma calls Ray3.2 a reasoning model that plans before rendering.^[2] Ray3.2 gives you 5- or 10-second clips from text, six aspect ratios (16:9, 9:16, 4:3, 3:4, 21:9 and 1:1) and an HDR pipeline. It gives you no audio.

The other text-to-video engines in the same workspace are third-party models billed in the same credits. Seedance 2.5 from ByteDance builds native audio in the same pass and, per Luma, reaches 30-second clips. Kling 4.0, the video model from Kling AI, runs from 3 to 30 seconds in Luma with multi-shot generation on by default. Veo 3.1, Google's video model, includes audio. MiniMax H3 renders 5 to 15 seconds at 2K and 24fps with matched sound. The credit rates differ more than the names suggest.

Every Ray3.2 row on Luma's pricing table also carries the note 'No keyframes: plus 30 credits'. Luma does not explain it further on that page, so read it there before you budget. On the older Ray3.14, which is still listed at the same prices, 10-second clips are text-to-video only and HDR is limited to 5-second clips at 720p or 1080p. MiniMax H3 is offered only to self-serve, non-enterprise accounts. Kling, Seedance, Veo and other outside models in Luma can change in availability, length limits and price, and Luma's limits may differ from other apps using the same model (review).

Step 2: Plan in Research Mode, then write a prompt Ray can plan from

Luma Agents, the workspace that grew out of the app once called Dream Machine, has two modes. Research Mode plans and Create Mode generates. Research Mode asks targeted questions, gathers references and writes editable shot lists or storyboard structures, but it produces no images, video or audio. Luma advises planning there first because it is cheaper than generating. For a text-to-video project, that is where the script, the scene order and the camera plan belong before a single credit goes to rendering.

Then write for the model. Describe camera movement in plain language. Luma's older Dream Machine API docs say camera motion is controlled by the words in the prompt, with occasional mismatches, and set prompts at 3 to 5,000 characters. You can attach optional references to a text prompt: an image, a storyboard, brand guidelines, an existing clip or a character reference. Keep one angle per reference image. Luma warns that sheets combining several angles can cause artifacts.

Watch for drift across versions. Luma's own guide says the Agent can hallucinate key elements from one version to the next unless you remind it to use your Master Reference Assets. Name them in the prompt each time. Also tell the Agent how many versions you want. By default it may generate several variations or retry automatically, which uses credits without asking.^[3] One sentence prevents that. Ask for one version.

Step 3: Draft at 360p, then render the final

The cheapest way to test a prompt is the Draft tier. A 5-second Ray3.2 draft at 360p costs 20 credits, against 100 at 720p and 400 at 1080p, all in SDR. Luma's own workflow for Ray3.2 goes in three moves: draft at low resolution, render the final at 1080p HDR, then export EXR frames. Its credit guide adds a second saving: upscale a strong draft instead of regenerating it, and switch off native audio on models that have it when you do not need sound.

By our arithmetic, Plus's 10,000 monthly credits cover 500 Ray3.2 drafts at 360p, 100 clips at 720p or 25 clips at 1080p, all 5 seconds long in SDR. Those counts assume every credit goes to that one job, with no failed or rejected takes. The same allowance covers about 25 five-second Kling 4.0 clips at 1080p, 14 from Veo 3.1 or 10 from Seedance 2.5 at 1080p.

Output format changes the bill as much as resolution. HDR costs 2x the SDR credits and HDR plus EXR costs 3x. By our arithmetic, that puts a 5-second 1080p Ray3.2 clip at 800 credits in HDR and 1,200 with EXR, so leave HDR off until the take is final. EXR export needs HDR switched on first. It attaches an EXR ZIP meant for grading and compositing tools such as Nuke, Resolve, Flame or Baselight.

Two billing rules shape how you iterate. Failed generations get their credits back, but takes you reject still cost credits (pricing). Unused credits expire at the end of each billing cycle, so a plan's allowance is use-it-or-lose-it. Drafting hard early in the month and rendering finals later is fine. Saving credits for next month is not possible.

Step 4: Give a silent Ray3.2 clip its sound

Ray3.2 output is silent by design, and Luma's Scenes storyboards confirm the Ray3.2 path has no audio. Sound gets added in the workspace. The Generate Audio panel has three categories: Speech, Music and Sound Effects, generated with ElevenLabs. On the pricing table, ElevenLabs v4 speech costs 21 credits per 1,000 characters, sound effects 25 credits per minute and music 98 credits per minute. Lip sync works by pairing the video with an audio track.

The alternative is to move the sound into the prompt. Luma's learning center says some of its models generate synchronized dialogue, ambient sound and music from a text description.^[4] Flux 3 from Black Forest Labs is one: its guide lists dialogue, ambience, effects, music and lip-synced speech inside the generation itself. Seedance 2.5 supports more than ten languages. If a shot needs a character speaking, write the line into the prompt and pick one of these models. If it needs a soundtrack under silent footage, Ray3.2 plus a music track is usually the cleaner route.

For short-form channels, Luma says its agents can write a script, split it into scenes, generate visuals, sync voiceover and format subtitles, then adapt one script into vertical cuts for Shorts, Reels and TikTok. Treat that as a workflow description, not a promise of finished edits. Each scene still costs its model's credits, and each voiceover costs speech credits.

When a text prompt is the wrong starting point

Text alone is the loosest brief you can hand a video model. Luma itself tells users to pick a starting strategy by the inputs they have: text, an image, existing video or keyframes. If a product, a face or a brand frame must look exact, start from a still and animate it; our image-to-video guide covers that route, including start and end frames. Ray3.2 accepts up to 16 keyframes per clip.

Length is the other limit. Ray3.2 text-to-video is priced only as 5- or 10-second clips. Longer stories need several clips joined, a Scenes storyboard, or a third-party model with a longer ceiling such as Kling 4.0 or Seedance 2.5. Resolution tops out at native 1080p; 4K comes only by upscaling. The Ray3.2 guide covers modes, Draft and HDR in more depth.

Access matters before any of this. Luma AI has no free plan on its pricing page; Luma says free trial credits are included on all plans without saying how many. The entry plan, Plus, is $30 a month billed monthly or $25 a month billed yearly for 10,000 credits a month. Commercial use comes with a paid plan and covers what you create while subscribed. Luma Agents is stated for the browser workspace.

The bottom line

Luma turns a prompt into a 5- or 10-second Ray3.2 clip at up to native 1080p, and the method that keeps it affordable is simple. Plan in Research Mode, ask for one version, draft a 5-second clip at 360p for 20 credits, then render a single final at 1080p and add HDR only when the take is approved. Ray3.2 is silent, so budget for ElevenLabs audio or write the prompt for a model with native sound (Seedance 2.5, Kling 4.0, Veo 3.1 or MiniMax H3) when the scene needs audio. Choose the model first. The prompt and the bill both follow from it.