What Ray3.2 Actually Does
Luma just released Ray3.2, and the headline is simple: you can now direct AI video frame by frame. Not just prompt and pray. Actually direct it.
Up to 16 keyframes per clip. Face and body performance tracking for up to eight characters simultaneously. EXR output that drops into post-production pipelines. A full API so you can build this into your own tools.
This is not another “AI video got slightly better at being consistent.” This is Luma building a tool for people who already know how to direct.
Multi-Keyframe Control
Place up to 16 keyframes in a single clip and the model fills in the motion between them. You set the narrative beats, the camera paths, the visual progression. The model executes. This is the difference between typing “car chase” and actually choreographing where the camera goes, when it cuts, and what the audience sees at each moment.
Performance Tracking for Faces and Bodies
The enhanced Performance Tracking carries over skeletal posture, gestures, and facial expressions from a reference performance. Track up to eight faces simultaneously in a single scene. The subtle head tilt, the way someone moves their mouth on a specific syllable, the gesture that sells a line, all preserved frame by frame.
EXR and HDR Output
Ray3.2 generates native HDR and exports 16-bit EXR. EXR is the format VFX and compositing pipelines run on. For the first time, an AI-generated video clip can go straight into DaVinci Resolve, Nuke, or After Effects without being crushed into an 8-bit colour space first. The dynamic range survives.
Surgical Edits, Not Full Reshoots
The Reframe tool lets you change aspect ratios, extend the frame, or replace backgrounds after generation, all while preserving the original lighting. Client wants a vertical crop for TikTok and a 16:9 for YouTube? One generation, two outputs, no re-prompting.
Longer Clips
Up to 20 seconds at 1080p. Long enough for a real scene with actual beats, not just a teaser loop.
The API
The full Ray3.2 control surface, keyframes, performance tracking, Reframe, HDR, is available as an API. No middleware. Studios can plug it into render farms. Agencies can build it into review pipelines. Platforms can embed it as a generation layer. This is the first time Luma has exposed the complete control surface via API, and it changes what you can build on top of it.
What This Unlocks
For anyone making video content, this changes the workflow in three ways:
1. Storyboarding becomes real. You can place keyframes that map to a real storyboard. Camera pans left here, cuts to close-up there, pulls wide for the establishing shot. The model respects your direction instead of inventing its own. If you have ever been frustrated by AI video that ignores half your prompt, this is the fix.
2. Post-production stays intact. EXR and HDR mean colourists and compositors get files they can actually work with. The AI clip becomes a production asset, not a social media curiosity. This is the difference between “cool demo” and “we can ship this.”
3. One generation, multiple deliverables. Reframe means you generate once and output at whatever aspect ratio you need. Horizontal, vertical, square. The lighting stays consistent. The background adjusts. This alone saves hours of rework for anyone delivering to multiple platforms.
How to Use It
Getting started with Ray3.2:
- Sign up at lumalabs.ai. Pro and Enterprise tiers get access to Ray3.2 features including the API.
- Start with keyframes. Upload or describe your first and last frame, then add intermediate keyframes for the beats you want to control. The model interpolates between them.
- Use performance tracking. Upload a reference video of the performance you want carried over. Ray3.2 maps the facial expressions and body movement onto your generated clip.
- Export for your pipeline. Choose EXR for compositing, HDR MP4 for review, standard MP4 for delivery. The format matches where the file is going next.
- Use Reframe for multi-platform. After generating your hero clip, use Reframe to create vertical, square, and wide crops without regenerating.
What to Watch Out For
The API is new. Luma built it for production pipelines but expect iteration. The 20-second clip limit is generous but still caps you at scene-level rather than full-act storytelling. Performance tracking works best with clean reference footage, shaky handheld input will not transfer as cleanly. And EXR file sizes are real: budget for storage if you are generating at volume.
The Bottom Line
Ray3.2 is the first AI video model that treats direction as a first-class input. You are not prompting. You are directing. Keyframes, performance tracking, EXR output, and an API that actually covers the full control surface. If you make video for a living, this is worth the time to learn.
Want more like this? Join MOKU Club for free. Weekly resources, early access to new guides, and occasional templates you can actually use. Join below.
