Luma Ray3.2 Lets Creators Direct AI Video Frame by Frame

Every AI video generator in 2026 works the same way. You type a prompt, click generate, and hope the model guesses what you meant. If the camera drifts left when you wanted it to hold still, or the actor raises the wrong hand at the wrong second, you regenerate and roll the dice again.

Luma’s Ray3.2, released on June 9, 2026, breaks that loop. It introduces up to 16 keyframes inside a single clip, letting creators dictate exactly what happens at specific moments throughout a 20-second sequence. The result is closer to directing than prompting, and for video creators who think in shot lists, this changes the production math entirely.

From Single Prompts to Keyframe Timelines

Previous AI video models, including Luma’s own Ray3 and Ray3.14, treated each generation as a single instruction. You wrote a text prompt, optionally included a reference image, and the model decided everything else: timing, camera movement, pacing, when subjects shift position, where the visual emphasis lands.

Ray3.2 replaces that approach with what Luma calls “frame-level control.” You place up to 16 keyframes across the timeline of a clip and specify what should happen at each point. The first keyframe might set a wide establishing shot. The third might pull to a close-up. The eighth might shift the subject’s posture. The model interpolates between your specified frames, maintaining continuity while following your direction.

For anyone who storyboards before they produce, this is significant. Instead of iterating through dozens of blind generations hoping one matches the shot list, you describe each beat and let the model connect them.

Eight Faces, One Coherent Scene

The other significant addition is simultaneous tracking of up to eight faces within a single scene. Ray3.2 maintains skeletal posture, hand gestures, and facial expressions across all tracked subjects, so a dialogue scene between two people or a group shot with multiple actors holds coherent performances throughout the clip.

Curious Refuge’s independent review scored the model at 7.7 out of 10, noting that motion quality (7.8) and prompt adherence (8.1) are the strongest areas. The reviewer found that Ray3.2 “rewards clear direction but punishes vagueness,” positioning it as a tool for creators who plan their shots rather than improvise with abstract prompts.

Professional Output That Fits Real Pipelines

Ray3.2 generates clips up to 20 seconds long at native 1080p resolution. For creators working in professional post-production pipelines, it outputs 16-bit EXR files with native HDR, meaning the footage can go directly into DaVinci Resolve, After Effects, or Nuke compositing sessions without dynamic range loss.

That’s a genuine differentiator. Most competing models output compressed MP4 files that lose highlight and shadow detail before the footage ever reaches a color grading suite. The EXR workflow matters for anyone compositing AI-generated elements into live-action footage or building visual effects shots.

The model also includes an enhanced reframing feature that adjusts aspect ratios and extends frames while preserving original lighting conditions. Creators repurposing horizontal content into vertical Shorts or Reels formats can reframe without regenerating entirely.

The Company Behind It

Luma AI is not a scrappy startup hoping this model lands. The company raised $900 million in a Series C round in November 2025 led by HUMAIN (Saudi Arabia’s Public Investment Fund), with participation from Andreessen Horowitz, AMD’s venture arm, Amplify Partners, and Matrix Partners. Post-money valuation sits at approximately $4 billion. Dream Machine, the platform that hosts Ray3.2, has over 30 million registered users and hit one million signups within 24 hours of its original launch.

Ray3.2 was developed in collaboration with creatives from the entertainment, advertising, and gaming industries, and the API is available for developers building custom production tools and render farm integrations.

Where Ray3.2 Struggles

No AI video model handles every scenario cleanly, and Ray3.2 has specific weak points worth knowing before you commit credits.

According to the Curious Refuge benchmark, the model struggles with animated or fantasy physics. Explosions, rapid particle effects, and impossible motion sequences break consistency. It also makes unexplained stylistic decisions, occasionally pulling footage into slow-motion without being prompted to. Objects in some sequences ignore physical laws entirely.

Temporal consistency scored 7.5 and visual fidelity scored 7.5, both solid but below the 8+ marks that would place Ray3.2 at the absolute top of the market. The reviewer positioned it “slightly below Seedance 2.0 and Kling 3.0 but comfortably among the stronger models on the market.”

For creators working in realistic, grounded scenarios (product demos, architectural walkthroughs, brand storytelling, talking-head variations) the model performs well. For fantastical or high-energy action sequences, Kling 3.0 or Runway Gen-4 may still deliver more predictable results.

Pricing and Credits

Luma runs a credit-based system through its Dream Machine platform. Current pricing breaks down across four tiers:

The free tier is useful for testing whether Ray3.2’s keyframe workflow fits your production style, but the draft-only resolution and watermark make it unsuitable for published content. The Plus tier at $29.99 is the realistic entry point for working creators who need commercial rights and clean output.

Two details worth knowing: credits do not roll over between months, and failed generations still consume them. With a model that rewards precise prompting, early experiments while you learn the keyframe system will burn through credits faster than experienced usage. Budget for a learning curve.

For iOS users, expect 25 to 30 percent higher pricing across all paid tiers due to Apple’s App Store commission. Purchasing through the web interface avoids that markup.

Where Ray3.2 Fits in Your Stack

The AI video generation market in mid-2026 is crowded. The way I’d frame the decision:

Choose Ray3.2 when you know exactly what each shot should contain. Narrative sequences, commercial work, product walkthroughs, brand stories with specific beats. The 16-keyframe system saves more production time than any marginal quality advantage from competing models, because you spend less time regenerating blind.

Choose something else when you need one good clip from a single prompt for social content. For that workflow, Kling 3.0 or Runway Gen-4 still deliver faster results with less setup. The complete 2026 AI video generation guide compares every major model’s strengths and pricing.

Combine them when your project mixes realistic dialogue scenes with VFX-heavy action. Generate the grounded footage in Ray3.2, the effects sequences in a model that handles unphysical motion better, and composite in Resolve or After Effects. The EXR output makes multi-source compositing straightforward.

Practical Tips for Getting Started

A few patterns from early testing:

Be specific at each keyframe. The model performs best when each keyframe includes concrete visual instructions: camera angle, subject position, lighting direction, action. “Man walks toward camera” produces worse results than “medium shot, man in navy suit walks toward camera, office corridor, overhead fluorescent lighting, even pace.”

Start with fewer keyframes. The maximum is 16, but using all 16 on your first attempt is like shooting a complex one-take with a new camera operator. Start with 4 to 6 keyframes, get comfortable with how the model interpolates between them, then add density from there.

Use the reference image system. Ray3.2 supports image-to-video with keyframe control combined. Starting with a reference image for your opening frame and placing keyframes for subsequent motion produces the most consistent results.

Reserve fantasy shots for other tools. If your project includes both realistic and fantastical sequences, play to each model’s strength. Ray3.2 handles grounded, well-lit footage with clear subjects. Let Kling or Seedance handle the physics-defying sequences.

Luma’s official Ray3.2 announcement includes API documentation for developers building keyframe control into custom production pipelines. For creators who want to test the keyframe workflow immediately, Ray3.2 is live on Dream Machine today.