Luma Ray

Overview

Ray3.2 is the video model Luma tells developers to integrate, on both the public Ray page and the API page fetched on 2026-10-16. Luma's own llm-info material in the same research pass described Ray3.2 as the current model and treated older Dream Machine and Ray2 names as deprecated. This page uses Ray3.2 as the current line and mentions Ray3.14 only where the pricing page still prints a credit table for it. There is no separate Luma tool page in the star set. The doors are Luma's site and API. This directory does not generate Ray clips.

Filter tags: text-to-video and image-to-video are true. The Ray page prices both, and the FAQ calls them text2video and image2video. Native audio is false. The FAQ answer "Do generations have native audio?" says text2video and image2video have no natively generated audio, so voiceover, music, and sound effects are a separate step. Modify Video and Reframe preserve original audio. That preservation is not generated speech. multi_shot is true because Multi-Keyframe is a documented control, up to 16 keyframes inside a single clip, including up to 16 keyframes for image-to-video. That is keyframe direction, not a Kling-style list of shot prompts. The filter uses the multi-shot bit because the information architecture treats multi-shot and multi-keyframe control as the same family of "more than one beat in a request." The body is the precise meaning. open_source is false.

Status: Ray3.2 is on public pricing, on the Ray marketing page, and on the API cost table. That is general availability for buyers of those plans or of pay-per-video API calls. A waitlist was not on those pages. A region lock was not on those pages. Team and Enterprise are "contact us," which is a sales motion, not a closed beta for the model. The API scale tier mentions dedicated capacity, a minimum of 4 units, an SLA, and a no-train guarantee. One unit there is defined as 1 request per minute on the base tier or 0.4 RPM on the max tier. That sentence is capacity packaging, not the per-video price.

Specs

Version line. Ray3.2 is current. Ray3.14 still has rows on the pricing page: text-to-video and image-to-video at 4 credits per second draft, 10 at 540p, 20 at 720p, and 80 at 1080p, with HDR at 4× on 3.14 versus 2× on 3.2. Use 3.14 only if you are maintaining an older integration. New selection should start at 3.2. The FAQ answer "What Is Ray3.2?" says: "Ray3.2 is our newest and most versatile video model. It powers four workflows in one: Modify Video, image2video, text2video, and Reframe." Dream Machine is not the current name to brief.

Ray page and FAQ, fetched 2026-10-16:

Field	Observed value
Resolutions	360p draft, 540p, 720p (default), 1080p.
Text and image length	5 or 10 second native clips.
Modify Video	Up to 20 seconds, depending on frame rate. 24 fps up to 20 seconds, 30 fps up to 15 seconds, 60 fps up to 7 seconds.
Reframe length	One FAQ answer says source videos up to 16 seconds. Another answer on the same page says up to 12 seconds. Both are printed. This page does not pick a winner.
Keyframes	Up to 16 in a clip, including image-to-video.
Aspects	9:16, 3:4, 1:1, 4:3, 16:9, 21:9.
Audio	No native audio on new text or image generations. Modify and Reframe keep source audio.
Container	Standard output mp4. Also native 16-bit HDR and 16-bit EXR (ACES2065-1 / AP0) for grading.
Modify inputs	mp4, mov, wmv.
Faces	Expressive facial performance for up to 8 faces, when that Modify feature is on. Lip sync preserved when face tracking is on, per the FAQ.

API page: "1080p output across the model. V2V at up to 20 seconds." HDR generation and 16-bit EXR export are repeated there. Draft 360p is on the credit table and is missing from the API dollar grid, which starts at 540p.

Credit table on the Ray page for Ray3.2, SDR base:

Task	360p draft	540p	720p	1080p
T2V/I2V 5s	20	50	100	400
T2V/I2V 10s	60	150	300	1200
V2V 5s	180	240	360	720
V2V 10s	360	480	720	1440
Reframe per second	10	20	40	120

HDR is 2× SDR credits. HDR plus EXR is 3×. The audio column on the pricing-page Ray3.2 block prints a dash, which matches the FAQ.

ElevenLabs is not a Ray audio feature. If a Luma screen offers an ElevenLabs connection, that is a separate product: /tools/elevenlabs. It was not a line on the Ray price table read here.

Prompt notes

The Ray page describes controls more than it gives a single canonical prompt. For Modify, the FAQ separates Motion (body, choreography, camera dynamics, which the tool tries to preserve) from Structure (shapes, edges, composition). A prompt that fights the source motion will not behave like a fresh text-to-video prompt. The page says Modify struggles with heavy camera shake, whip-fast action, blur, darkness, low resolution, crowded overlapping subjects, and occluded faces. That limitation is the vendor's own FAQ, not a lab note from this directory. No clip was generated here.

Arena scores were not opened. Not observed.

Where to run

1. The Ray product page: lumalabs.ai/ray. Specs, keyframes, credit table, and the FAQ JSON that answers the audio question.
2. The API cost page: lumalabs.ai/api. Pay-per-video SDR dollars, HDR multipliers, and the statement that there is no minimum commitment on that pay-per-video tier.
3. Plan cards: lumalabs.ai/pricing. Plus, Pro, and Ultra credit pools, plus the same Ray3.2 credit math and rows for other companies' models.

The API page says partners include a path such as Fal. A fal Ray3.2 llms.txt was not saved in this pass, so a fal dollar card for Ray is not observed. Do not invent one. Luma's own API table is the dollar channel.

Multi-channel pricing

Two Luma channels were opened on 2026-10-16: subscription credits and API dollars. They are both official. A third-party host price was not observed.

API, SDR, pay per video, no minimum commitment. Text-to-video and image-to-video, 5 seconds: $0.15 at 540p, $0.30 at 720p, $1.20 at 1080p. Ten seconds: $0.45, $0.90, and $3.60. Video-to-video, 5 seconds: $0.72, $1.44, and $2.16. Ten seconds: $1.08, $2.16, and $4.32. Reframe per second: $0.06, $0.12, and $0.36 at 540p, 720p, and 1080p. The page says prices shown are SDR. HDR is 2× those figures. HDR plus EXR is 3×. A 1080p 5-second HDR text clip is $1.20 times 2 = $2.40 derived. Draft 360p has credits and does not have a row in this dollar grid. Rate limits apply and that pay-per-video tier has no latency SLA, per the page. The scale tier is contact sales.

App credits, Ray page and pricing page. The table in Specs. Plans that buy the credits: Plus is $30 per month or $300 billed yearly, 10,000 credits, commercial use, Luma and third-party models. Pro is $90 per month or $900 billed yearly, 40,000 credits, and 4× Luma Agents usage. Ultra is $300 per month or $3,000 billed yearly, 150,000 credits, and 15× agents. The pricing metadata says save up to 20% with yearly. The billed-yearly totals are not twelve times the monthly price: Plus $300 versus $30 × 12 = $360, Pro $900 versus $90 × 12 = $1,080, and Ultra $3,000 versus $300 × 12 = $3,600. Each is about 17% less. The plan data on that page also includes yearlyPrice values of $25, $75, and $250, which are those yearly totals divided by 12. None of the three is a full 20% off. The "up to 20%" line is the vendor headline. No free plan was on the cards that were extracted. Team and Enterprise are contact.

A worked credit check, derived: one 720p 5-second text clip is 100 credits. Plus's 10,000 credits cover 100 such clips if nothing else is spent and HDR is off. One 1080p 5-second text clip is 400 credits, so the same pool covers 25. A 1080p HDR clip is 800 credits. These products assume the table and no failed-generation rebate. A rebate rule was not in the extracted FAQ answers.

Third-party models on the same pricing page, including Seedance, Kling, Veo, and MiniMax, spend the same credit pool and are not Ray. Their rows belong on their own model pages. Mixing a Ray 100-credit clip with a Seedance per-second credit row is how teams blow the pool.

Workflow and platform fit

Labels only. Real workflow guides for these names are not linked.

Faceless. Ray3.2 can do 9:16 and a 5 or 10 second native clip, which is a B-roll beat. It will not speak. Every faceless cut needs an external voice, then captions. Use /tools/elevenlabs and /tools/capcut. Modify can restyle footage you already shot and keep its audio, which is a different faceless pattern: film once, restyle many times, keep the take's sound.

UGC ads. Product swap and environment change are named on the Ray page as workflows. They are marketing names for Modify-style jobs, not a separate price SKU in the table. A testimonial face that must stay locked is a weaker pitch than HeyGen unless your test shows Modify's face tracking holds. The vendor says face tracking preserves lip sync. It also lists occluded faces as a failure mode. /tools/creatify remains the template tool for a structured UGC ad.

AI avatar. Ray is not an avatar engine. Up to 8 faces in Modify is performance transfer, not a photo-to-spokesperson API. See /models/heygen-video.

Long-to-short. Modify's maximum is on the order of 7 to 20 seconds depending on frame rate, and reframe is 12 or 16 seconds depending on which FAQ line you cite. A long episode must be cut first. Ray does not replace a long-to-short editor.

Captions and cover. No caption model is in the Ray spec. A cover can be a frame from the mp4 or a still from Uni-1, which the API page sells as a separate image model. Uni-1 prices were on the API page and are out of scope for this video sheet beyond that mention.

Voice and publish. Publish path is silent picture plus a voice track, or Modify of a clip that already has a voice. 9:16 is in the aspect list, so Shorts, Reels, and TikTok frames are specified. HDR and EXR are for a grade in Resolve or Nuke, which the API page names. They are the wrong master to upload untouched to a phone feed. Export or transcode to a normal mp4 before you schedule the Short.

Shorts vertical fit. Strong on aspect and on the existence of a draft 360p credit tier for cheap previz. Weak on audio and on length if the Short needs a continuous 20-second generated scene from text. Text and image generations stop at 10 seconds. A vertical ad cut from a modified source can run longer, up to the frame-rate cap, and can keep the performer's audio. That is the Shorts pattern Ray actually documents.

Pros

• The audio question has a direct FAQ answer, so you do not have to infer silence from a dash in a table. The dash and the answer agree.
• Sixteen keyframes is a concrete control, repeated for image-to-video and for Modify.
• 9:16 is in the aspect list, alongside a cinematic 21:9, so one model card covers a Short and a wide master.
• HDR 2× and HDR+EXR 3× are written next to the credit table, and the API repeats the same multipliers in dollars.
• App credits and API dollars were both public on the same day, so a studio can compare a subscription pool with pay-per-video without a sales call.
• Modify's failure modes are published by the vendor, which is more useful than a feature grid that only lists wins.

Cons

• New text and image clips have no native audio. Any workflow that wanted one-request speech should look at Kling, Seedance, or Veo instead.
• Text and image generations are only 5 or 10 seconds. The longer numbers belong to Modify and depend on source frame rate.
• Reframe's maximum length is stated as both 16 seconds and 12 seconds on the same page. Quote both until Luma edits one of them.
• Draft 360p is missing from the API dollar grid. A developer cannot assume the 20-credit draft maps to a published dollar.
• 1080p text-to-video is $1.20 for 5 seconds SDR, before HDR. That is a different budget from 540p at $0.15. The cheap headline and the master price should not be swapped.
• Ray3.14 rows are still on the pricing page. An old integration can bill the 4× HDR multiplier by mistake.
• No non-Luma price card was saved. Host arbitrage for Ray is not observed.

Alternatives

• Seedance generates audio inside the fal price and runs up to 30 seconds on the 2.5 cards. Ray is the one with keyframes, modify, and EXR.
• Kling has native audio, a multi-shot prompt list, and an official unit table. It does not publish a 16-keyframe Modify FAQ of the kind Ray does.
• Veo has sound generation, an 8-second cap, and a GA card in us-central1. Ray's pay-per-video API is the easier public dollar meter. Veo's is split between preview Gemini prices and a Vertex card without pay-as-you-go.
• Runway Gen is 12 Runway credits per second with an HDR surcharge and no confirmed 9:16 enum on its price guide. Ray's 9:16 is explicit.
• MiniMax Hailuo is priced per second on MiniMax's own pay-as-you-go page, without Ray's keyframe and EXR story.

Related tools

• /tools/elevenlabs is the voice step the Ray FAQ tells you to add.
• /tools/capcut holds captions, covers, and the cut that brings Ray beats under one timeline.
• /tools/runway, /tools/kling, and /tools/seedance are the other studios when you leave Luma's credit pool.
• /tools/higgsfield hosts several of those alternative models. A Ray endpoint was not in the Higgsfield catalog snapshot read for this pass.
• /tools/heygen and /tools/creatify cover presenters and ad templates that Ray's Modify path does not productize as an avatar.

Luma's own app is the primary tool and it does not have a /tools slug in this directory yet. The three links above satisfy the related-tool rule. The missing Luma tool page is a directory gap, not a claim that Ray has no app.

FAQ

Does Ray3.2 generate audio?

Not for text-to-video or image-to-video. The official FAQ says there is no natively generated audio on those modes. Modify and Reframe keep the audio that was already in the source file.

How long is a clip?

Text-to-video and image-to-video: 5 or 10 seconds. Modify: up to 20 seconds at 24 fps, 15 at 30 fps, and 7 at 60 fps. Reframe: the page says both 12 and 16 seconds in different answers.

How many keyframes?

Up to 16, including image-to-video, according to the FAQ.

What does a 1080p clip cost on the API?

A 5-second SDR text or image clip at 1080p is $1.20 on the API table fetched 2026-10-16. HDR doubles that. HDR plus EXR triples it. The same clip in the app is 400 credits before multipliers.

Is Ray open source?

No weight release was on the pages opened. open_source is false. The API is a paid service with a separate scale tier for an SLA.