Research profile · Ray 3 (the benchmarked release); Ray 3.2 of June 2026 is current
Luma Dream Machine, researched.
Documented output specifications well beyond the rest of this set. Provisional findings. Explore the findings, limitations and 5 cited sources below.
Luma Ray 3 is supported for delivery into a graded pipeline — HDR and 16-bit EXR output are documented further than anything else in this set — and places mid-table on blind preference. What we know about its weaknesses comes from Luma itself: no audio, character identity that can drift between generations, and abrupt transitions on extend and loop. Seven of eleven criteria rest on the vendor's own documentation, so the verdict is Limited.
- Researched scope
- Ray 3 (the benchmarked release); Ray 3.2 of June 2026 is current
- Best supported for
- Delivery into a graded pipeline, where HDR and EXR output matter
- Principal limitation
- No audio; identity drift between generations; most findings are the vendor's own admissions
- Accepted sources
- 5 · 7 of 11 criteria supported
Best for, and not for
Best supported use
Delivery into a graded pipeline, where HDR and EXR output matter. The strongest finding is documented output specifications well beyond the rest of this set.
Where the evidence warns
No audio; identity drift between generations; most findings are the vendor's own admissions.
What the evidence supports, criterion by criterion
7 of 11 applicable criteria have enough accepted evidence to say something. The rest are listed below them, unscored.
| Criterion | Evidence | What we found |
|---|---|---|
| Visual fidelity and aesthetics | Limited evidence | Ray 3 places mid-table on blind pairwise preference in the silent text-to-video arena: Elo 1189 from 5,476 votes.Scoped to Ray 3 because that is the release independently measured. Ray 3.2, a June 2026 point release, is current and unbenchmarked; the page says so. Preference Elo speaks to overall fidelity only. |
| Temporal stability and artifact control | Limited evidence | Luma's FAQ states that 'legacy features like Extend, Loop and Keyframe Interpolation occasionally have abrupt transitions.'A workflow-transition limitation admitted by the vendor, not a measured flicker rate on base generation. The weakest of the three admissions as evidence for this criterion. |
| Workflow and editability | Limited evidence | Documented workflow beyond plain generation: Ray3 Modify for altering existing footage, video-to-video at up to 20 seconds on Ray3.2, and keyframe direction.Feature documentation only; no accepted evidence on how well these behave in practice. |
| Output specifications and delivery | Moderately supported | Ray3.2, released June 2026, is documented at 1080p across the model with video-to-video up to 20 seconds. Ray3 supports 540p at 5-10s and 720p/1080p at 5-20s, with native HDR and 16-bit EXR export.EXR and HDR output is unusual in this set and is documented rather than inferred. |
| Pricing and value | Moderately supported | Published tiers: Lite $9.99/month for 3,200 credits, Plus $29.99 for 10,000 credits with commercial rights and no watermark, Unlimited $94.99 adding unlimited relaxed generations. Ray3 1080p SDR is 330 credits for 5 seconds.iOS pricing is higher than web on every tier; the figures above are the web prices. |
| Character and object consistency | Limited evidence | Luma's FAQ admits that 'character consistency can vary between generations with potential shifts in age or gender.'A vendor-stated limitation without a rate. It speaks to consistency across generations, not within one clip, which is the weaker half of the criterion. |
| Native audio and lip sync | Limited evidence | Ray 3 does not generate audio. Luma's own FAQ answers the question directly: 'No, audio is currently not supported.'A documented absence of a capability, which answers the criterion. Ray 3.2 may differ; this is the Ray 3 release we scope. |
Not established
Researched, but the available public evidence does not support a conclusion. These are gaps, not low scores.
- Prompt or script adherence and control — Searched the VBench and VBench-2.0 leaderboards on 13 September 2026 by full model-column inspection. No version matching our scope appears in them.
- Camera motion — Searched the VBench and VBench-2.0 leaderboards on 13 September 2026 by full model-column inspection. No version matching our scope appears in them.
- Physics and material realism — Searched the VBench and VBench-2.0 leaderboards on 13 September 2026 by full model-column inspection. No version matching our scope appears in them.
- Text rendering — Searched Luma's Ray 3 FAQ and launch material on 14 September 2026; nothing on text rendering.
Pricing
Published list price, normalised to a five-second generation. Assumes full use of the plan allocation, before tax, no retakes — so real cost per clip you keep is higher.
| Plan | Cost / 5s | Resolution | Audio |
|---|
All Luma Dream Machine pricing scenarios and how they were normalised. Prices verified 2026-09-14 and change without notice.
Alternatives
Same category, so genuinely comparable. We do not rank across categories.
- Kling AI — Complex physical movement and holding a character's face steady across a shot. Moderately supported.
- Higgsfield — Camera-move control and stylised cinematic shots, with access to several models in one place. Not established.
- Runway Gen-4.5 — Creative and editorial work — generative editing, motion brush and in-context controls. Moderately supported.
Sources
Every finding above traces to one of these. Version match records how closely each source matches Ray 3 (the benchmarked release); Ray 3.2 of June 2026 is current.
| Source | Type | Version match | Accessed |
|---|---|---|---|
| Luma Labs | Official documentation | Exact | 2026-09-12 |
| Luma Labs | Official documentation | Exact | 2026-09-12 |
| Luma Labs | Official documentation | Exact | 2026-09-12 |
| Artificial Analysis | Independent benchmark | Close | 2026-09-13 |
| Luma AI (Celina B.) | Official documentation | Exact | 2026-09-14 |