AI video generation just learned a new word: "direction."
Until now, AI video tools let you set a start frame and an end frame at best — everything in between still had to be patched together by hand in an editor. Luma AI's new Ray3.2 model breaks that ceiling by letting you place up to 16 keyframes in a single clip.
Stuck with clips that
"kind of" looked right
For the past few years, AI video generation moved from text prompts to, at best, a start and end image. Runway went a long time without a native end-frame input at all, and Kling 3.0 only recently added start/end two-point control. Getting the exact cut you wanted still meant pulling generated footage into an editor and reassembling it by hand.
Luma AI's Ray3.2, released June 9, 2026, changes that premise. A single clip can now hold up to 16 keyframes, each one setting camera angle and composition directly, with the model interpolating the motion in between. It's the shift from a tool that "kind of" shoots what you meant to one you can actually direct.
| Prior leading tools | Ray3.2 |
|---|---|
| Runway has no native end-frame input | Up to 16 keyframes per clip |
| Kling 3.0 caps out at start/end, 2 points | Camera path and framing set at each point |
| Mid-clip cuts fixed by hand in an editor | Motion between points interpolated automatically |
| In-app generation only, hard to integrate | Ray3.2 ships as an API for the first time |
Ray3.2 by the numbers
Source: Luma's official API page, "Build with Luma APIs," and Luma's official announcement (June 9, 2026).
Place the points,
let the model fill the gaps
The mechanics are simple. Drop up to 16 anchors on the timeline, each describing what the shot should look like at that moment, and Ray3.2 interpolates the motion between them into one continuous clip.
Place the anchors
At any point on the timeline, set up to 16 keyframes specifying composition, camera position, and subject pose.
The model interpolates
Ray3.2 smoothly generates the motion between adjacent keyframes. Denser keyframes get you closer to the exact camera move you had in mind.
Export as HDR/EXR
Output supports native HDR generation and can be exported as 16-bit EXR, feeding AI-generated footage directly into professional color-grading tools like DaVinci Resolve.
From writing a prompt
to directing a timeline.
Who it helps, and how
Video producers & designers
You can match a storyboard's shot list without post-hoc fixes. The trade-off: the more keyframes you use, the closer you get to what you intended, but designing the point layout itself takes practice.
Marketers
Ad and social clips can hit the exact composition you wanted at the exact moment. That said, it's overkill to design 16 points for a five-second vertical clip — the payoff scales with the length of the piece.
Engineers
Ray3.2 is available as an API for the first time, so it can be embedded directly into your own product. Work that previously had to go through Luma's app can now be automated as a feature of your own SaaS.
Not without caveats
With Ray3.2, the achievable precision of AI video control jumped from a two-point start/end pair to as many as 16 points — that much is confirmed. But this is one move in a race where Sora, Kling, and Veo are all pushing toward the same goal: getting past prompt-only control into frame-level direction. The number 16 shouldn't be read in isolation from that competition.
Practical limits remain, too. Designing 16 keyframes is meaningfully more work than typing a single prompt. Generation is also billed on a credit basis that scales with resolution and duration, so longer, higher-resolution projects could cost more than expected. This isn't "the editor becomes unnecessary" — it's closer to "part of editing moves earlier, into timeline design."