Some links below are affiliate links; buying through them supports the site at no cost to you.
Luma ships two model families of its own in October 2026: Ray for video and Uni for images. Ray3.2, released June 9, 2026, is the one that matters to an editor, because it is the only major video generator that publishes an EXR export: a zip of 16-bit frames in ACES2065-1 instead of a finished mp4. Ray3.14 is the cheaper, faster model that still does the extend and loop jobs Ray3.2 dropped. Modify Video is the video-to-video mode that runs on real footage. Everything else Luma once sold, the Photon image model, the Genie text-to-3D model and the phone app that captured Gaussian splats, is renamed, retired or frozen.
Last updated: October 2026. Models, specs and terms checked October 29, 2026, on lumalabs.ai, its Learning Center, the API page and the Terms of Service. Luma has no affiliate program for publishers, so no link on this page pays us. Luma revises its model list every few months; a dated line here records each change.
TLDR
Use Ray3.2 for any shot that has to survive a grade or a comp: it generates native 16-bit HDR, exports EXR, holds up to 16 keyframes in one clip, and its Modify Video mode keeps an actor’s performance while you change everything around it. Use Ray3.14 when you need to extend a clip, loop it, or burn through drafts at a third of the cost. Use Uni-1 for the still frames you feed either model as keyframes and references. Ignore the Photon and Genie names in older tutorials; neither is on the current pricing page, and Genie was shut down on January 1, 2026. Every model is on the $30 Plus plan, so the choice is never which tier unlocks it, only how many credits a job burns.
Luma’s models in October 2026
| Model | Job | Status | Editor-relevant spec |
|---|---|---|---|
| Ray3.2 | Text-to-video, image-to-video, Modify Video, Reframe | Current, web app and API | 360p to 1080p, 5 or 10 s native, Modify up to 20 s, 16-bit HDR, EXR export, no native audio |
| Ray3.14 | Text-to-video, image-to-video, Modify, Extend | Current, web app | Native 1080p, 3x cheaper than Ray3 at 720p, Modify up to 18 s, HDR but no EXR, no character references |
| Ray3 | Video generation | Superseded | September 2025 model; still named in Luma’s help pages and as an API fallback |
| Ray2, Ray2 Flash | Video generation | Legacy | Early 2025; some Reframe integrations still default to it |
| Uni-1, Uni-1.1 | Image create and modify, Layers | Current, web app and API | Up to 9 role-labelled references, seeds, nine aspect ratios, 30 credits an image |
| Photon | Image generation | Legacy | 2024 image model; priced only on a December 2025 help page |
| Genie | Text-to-3D | Sunset January 1, 2026 | Exports still allowed; lumalabs.ai/genie now redirects to the homepage |
| Luma 3D Capture | NeRF and Gaussian splat capture | Frozen, iOS only | App Store version 1.3.14 from January 2026; web entry points gone |
The product called Dream Machine in 2024 and 2025 is now just Luma, or Luma Agents inside the app, and the agent layer routes to third-party models as well: Seedance, Kling, Veo, MiniMax, Flux, GPT Image, Nano Banana, Seedream, Ideogram, ElevenLabs and Sync Lipsync all appear on the pricing page with their own credit rates. This guide covers Luma’s own models. The third-party rows are the same models you can buy elsewhere, at Luma’s exchange rate.
Ray3.2: the model with an EXR export
Ray3.2 generates at 360p, 540p, 720p and 1080p, and the first thing to know is what the resolutions are for. Luma’s own training article tells you to explore at 540p in Speed mode, review at 720p, and render 1080p in Quality mode only for the shot that ships. Text-to-video and image-to-video produce 5- or 10-second native clips. Modify Video returns a clip the same length as the source up to 20 seconds, with the ceiling set by frame rate: 20 seconds at 24 fps, 15 at 30 fps, 7 at 60 fps. Reframe takes a source of 12 to 16 seconds, depending on which of Luma’s two FAQ answers you read, and outputs 9:16, 3:4, 1:1, 4:3, 16:9 or 21:9.
Keyframes are what changed from Ray3. Earlier Ray models took a start frame and sometimes an end frame and filled in the rest. Ray3.2 takes up to 16 keyframes at chosen points in the clip at launch, and Luma’s video-to-video guide now says 64 at arbitrary source-frame indexes for Modify. Either way the workflow is the same: export frames from a source, paint or regenerate them in Uni-1 or Photoshop, drop them back at the same frame numbers, and the model interpolates the look between them. Performance Tracking holds skeletal pose for up to eight bodies and Expressive Facial Performance carries expression for up to eight faces.
The output formats are what set Ray3.2 apart. Standard output is an mp4. Switch HDR on and Ray3.2 generates “native 16-bit color,” at twice the credits. Switch EXR on, which requires HDR, and the generation arrives with “an EXR ZIP variant” attached, at three times the credits. Luma’s colour-space article, written for the Ray3.14 HDR model in March 2026, gives the spec that the Ray3.2 pages inherit: OpenEXR, 16-bit half float, DWAB compression at quality 45, ACES2065-1 with AP0 primaries, scene-referred linear, no tone mapping. In Luma’s words, “You’re not getting final pixels, you’re getting a source asset with headroom.” Audio: text-to-video and image-to-video generate none, and Modify and Reframe pass the source track through.
Ray3.14: the fast one, and the only one that extends
Ray3.14 arrived January 26, 2026, with “native 1080p across core Dream Machine workflows,” 4x the speed of Ray3 at 720p and a third of the cost. On the October pricing page its text-to-video and image-to-video rates match Ray3.2’s, so the saving against Ray3.2 is in the jobs Ray3.2 does not do. Extend Video is a Ray3.14 feature, priced at the standard SDR rate per 5-second block. Loops are a Ray3.14 feature. Changing a clip’s duration in Modify is a Ray3.14 feature; Ray3.2’s Modify always returns the source length. Luma’s own guide: “Use Ray 3.14 Modify only when you specifically need to change duration or create a loop.”
The limits are published with the launch. Character references are not supported in Ray3.14. HDR is available at twice the credits, but “HDR/EXR in Modify Video isn’t supported,” and on the October 2026 pricing page Ray3.14 has no EXR row at all. So for a plate that has to go to a colourist, the model is Ray3.2.
Modify Video: video-to-video on real footage
Modify Video is where Luma’s models differ from a text prompt box. You give it a clip you already shot, and it returns the same clip with a new look, wardrobe, environment, material or lighting, keeping the camera move and the performance. The current version runs on Ray3.2 and is reached by right-clicking a video on the canvas and choosing Modify Video, which Luma calls “the most common and recommended approach.” Inputs are mp4, mov or wmv.
The controls are two sliders and three locks. Motion, 1 to 9, decides how much of the source’s movement carries over; at 1 it mostly holds the camera move, at 9 it tracks small movement across the frame. Structure, 1 to 9, decides how tightly shapes are held; Luma describes 8 as the setting that holds the subject and frees the background, and 9 as locking every edge in the frame. Faces, Bodies and the Poses or Blocking skeletal modes are the character locks, and the guide’s advice is to run the lightest setup first and add a lock only to fix what drifts. Lip sync survives “when Face tracking is on.” Footage that works is steady, well lit and in focus with one clear subject; heavy shake, whip pans, crowds and occluded faces are the published failure cases. The Enhance toggle rewrites your prompt for video-to-video and is on by default, so switch it off once a prompt is approved.
The prompt itself describes the finished frame, not the change. Luma’s examples: not “change the sky to purple” but “a purple sky over the mountain landscape,” and end with “Preserve all other elements including subject identity, pose, and lighting.” Cost is per second of source, 36 to 144 credits at SDR depending on resolution, so a 20-second 1080p clip is 2,880 credits before HDR. Trim the source before you upload it.
Uni-1: the image model that feeds the keyframes
Uni-1 replaced Photon as Luma’s image model in early 2026, and Uni-1.1 is the API version from May. Luma calls it “a multimodal reasoning model that can generate pixels,” which in practice means two modes: Create Image, which makes a new composition from a prompt and up to nine reference images, and Modify Image, which edits a source image and keeps its dimensions. The references are the useful part. Each one takes a role in the prompt, “Use IMAGE1 as a STYLE reference,” “IMAGE2 as LIGHTING,” and the guide’s rule is that an unlabelled reference is a guess. Seeds make a result repeatable, prompts run to 6,000 characters including JSON, and Layers, added July 2026, splits a result into editable layers.
For an editor the job is keyframes. Generate the look of a shot as stills, lock the seed, then hand those stills to Ray3.2 at the frame numbers where the look should land. Thirty credits an image in the app, about nine cents through the API.
Photon, Genie and 3D capture: what Luma retired
Photon was the image model from December 2024. It is priced on Luma’s credit-system help page, dated December 15, 2025, at four credits an image, and it is absent from the pricing page, the model picker copy and the API page in October 2026. A Photon tutorial describes the Uni-1 workflow with older screenshots.
Genie was the text-to-3D model, free in 2023 and folded into the capture app. The App Store notes for Luma 3D Capture version 1.3.14 say it directly: “We’ve officially Sunset Genie since January 1, 2026. Thank you for being part of the journey. You can still export your content.” The Genie page on lumalabs.ai redirects to the homepage.
Luma 3D Capture, the iPhone app that built NeRFs and Gaussian splats and started the company, is still on the App Store as a free download rated 4.5 from 1.9K ratings, and still exports. Its web entry point returns a 404, its Unreal plugin on Fab was last updated October 15, 2024 and supports Unreal 5.1 to 5.3, and no 2026 page on lumalabs.ai mentions it.
Which plan runs which model
All of them, on any paid plan. Plus at $30 a month, Pro at $90 and Ultra at $300 differ by credits, 10,000, 40,000 and 150,000, and by a multiplier on how much the Agent will do unattended. There is no tier that unlocks Ray3.2, HDR or EXR; there is only the credit bill, which triples for an EXR render. The free sign-up has no published allowance and its output is non-commercial under the May 2026 terms.
Getting a Ray3.2 EXR into Resolve
The zip unpacks to a numbered EXR sequence. In Resolve, set the project to ACES (ACEScct working space is fine) and tag the sequence’s input colour space as ACES AP0, or in a Resolve Color Management project set the clip’s input to ACES2065-1. The frames are scene-linear and carry values above diffuse white, so a Rec.709 output transform will show you something sensible at once and a Rec.2020 PQ transform will show you the HDR headroom the model generated. Do not apply a camera LUT; there was no camera. Treat the frame rate as the clip’s native rate and conform it on the timeline rather than retiming. Premiere imports the same files as an image sequence and takes the input colour space from the project’s colour management; if you only need an offline, the mp4 that came with the generation is a display-referred version of the same clip.
If the plate is going to a comp rather than a grade, the generative footage in Premiere and Resolve guide walks through the round trip, and the small-studio VFX stack lists the hardware that plays a 1080p EXR sequence without a cache.
FAQ
Is Ray3.2 the same as Dream Machine?
Dream Machine was the name of Luma’s app. The app is now called Luma, with Luma Agents as the workspace, and Ray3.2 is the current video model inside it. Ray models are what Dream Machine always ran.
Does Luma Ray3.2 generate audio?
No. Text-to-video and image-to-video clips are silent. Modify Video and Reframe keep the source audio. Voice, music and lip sync in the app come from third-party models billed at their own credit rates.
Can Luma export EXR files?
Yes, from Ray3.2 in the web app and the API. EXR requires HDR to be on and costs three times the SDR credits. The frames are 16-bit half float in ACES2065-1. Ray3.14 generates HDR but not EXR.
Which Luma model is best for video to video?
Ray3.2, unless you need to change the clip’s length or make a loop, which only Ray3.14’s Modify does. Ray3.2 keeps the source duration, holds up to 20 seconds, and adds multi-keyframe control, face and body locks, and EXR output.
Is Luma Genie still available?
No. Luma sunset Genie on January 1, 2026, and the Genie page now redirects to the homepage. Existing content can still be exported from the Luma 3D Capture app.
Does Ray3.2 work inside Premiere Pro or Resolve?
Not as a panel. Ray models have shipped inside Adobe Firefly since 2025, and Luma’s app has connectors for Frame.io and Figma, but there is no Luma extension for Premiere or Resolve. You generate in the app or through the API and import the mp4 or the EXR sequence.
In conclusion
Luma’s lineup in late 2026 is two models wide. Ray3.2 is the one built for post, with the EXR path and the keyframe control that no other generator publishes, and Ray3.14 is the cheaper tool for extends, loops and drafts. Uni-1 makes the stills that steer them. The names you will meet in older guides, Dream Machine, Photon and Genie, are a rename, a replacement and a shutdown. Pick the model by the deliverable: mp4 for a social cut, Ray3.14 for anything that needs to run longer than ten seconds, Ray3.2 with HDR and EXR for anything a colourist will touch.
