Luma Ray 3.2, the video AI people pick for its cinematic look

A Video AI Built Alongside Filmmakers

Most video AI models compete over who can make clips longer or add sound. Luma Labs went the other way. The company developed Ray 3.2, released on June 9, 2026, together with people from film, advertising, and game development, and its main bet is on control and image quality.

Luma frames it like this, with the new version users are no longer just writing prompts, they're directing. The model is supposed to follow the creator rather than hand back a random pretty result that can't be reproduced.

That matters for regular users too. If you want a clip that looks like real footage rather than typical AI video, Ray 3.2 is worth trying early on.

What Luma Ray 3.2 can do

Ray 3.2 covers two main scenarios. You can describe a scene in text and the model shoots a clip from scratch. Or you can attach a photo, and the model animates it, with the picture becoming the first frame of the video.

Finished clips run 5 or 10 seconds at up to 1080p. There are six aspect ratios, from vertical 9 by 16 for phones to widescreen 21 by 9, so there's a fit for any platform.

Third-party reviews note that the model plans the logic of a scene before it draws the motion. That's why clips come out without the jitter and drifting details that usually give away AI video. It shows most in product reveals, atmospheric background scenes, and animating finished illustrations.

Why a video AI needs HDR

This is the main thing that sets Ray 3.2 apart from competitors. The model can generate video directly in HDR with 16-bit color, and export frames in the EXR format used in film and advertising.

In plain terms, regular video stores a limited set of shades, and if you push the colors hard later, the image breaks down with banding and muddy shadows. 16-bit video has a huge reserve of shades, so you can push it in color grading as boldly as footage from a professional camera.

Who actually needs this,

• People who edit clips and fine-tune color in tools like DaVinci Resolve

• Advertising teams who need to match the image to a brand's colors

• Anyone blending AI shots with real footage who wants the seam to disappear

If you're just making a clip for social media and don't plan to touch the color, you don't need HDR. The regular mode looks great and costs half as much.

Frame-by-frame direction and working with existing video

The second major addition concerns control over motion. Earlier Ray models took a first frame, sometimes a last one, and decided on their own what happened in between. Ray 3.2 lets you place up to 16 keyframes within a single clip, so you decide where the image shifts, where a movement lands, and where the key moment of the scene hits.

A separate direction is reworking existing video. The model takes a source clip and changes the background, wardrobe, setting, or lighting while keeping the actors' movement, facial expressions, and timing. Luma says the model tracks body poses and facial expressions for up to eight people in frame, and a reworked clip can run up to 20 seconds at 1080p.

These features are aimed primarily at studios and agencies. For most users, the main tool remains generation from text and photos, which is what we'll cover next.

How to use Luma Ray 3.2 on unitool.ai

In our interface there's nothing to switch, the mode is picked automatically. Type only text and Ray 3.2 shoots a clip from scratch. Attach a photo with the paperclip and the model animates your picture instead, making it the first frame.

Here are two examples. From text,

Camera slowly tracks along an empty cafe terrace at dusk, string lights sway, 
reflections shimmer on wet cobblestones

And if you attach a photo of that same terrace, the prompt only describes what should start moving,

String lights sway in the wind, rain starts to fall on the tables, 
subtle camera push in

Notice the difference. In the second case there's no need to describe the terrace itself, it's already in the photo. The prompt handles only motion, weather, and the camera.

How to describe a scene so it works the first time

Ray 3.2 has a simple order in which it understands a request best. First describe who's in frame and what they're doing, then how the camera behaves, and finally the lighting and time of day.

The main rule is one camera move per clip. If you ask for a push in, a pan, and an orbit all at once, the model starts to thrash, while a single move gives a noticeably cleaner result. Ray understands directing terms like push in, pull out, pan, handheld, and top-down.

Here are ready-made prompts you can copy and adapt,

A woman in a yellow raincoat walks along a wet embankment, the camera follows her from the left, late evening, lights reflecting in puddles

Coffee pours into a glass cup, macro shot, steam rising, soft morning light from a window, static camera

A drone rises over a misty pine forest, the first rays of sunlight breaking through the trunks

The model understands Russian, but English remains the most predictable option. If a result in another language misses the mark, run the same prompt in English and the difference is often obvious right away.

Settings and how to avoid overpaying

The right-hand panel has only a few settings, but they drive the price heavily. Let's go through each.

Quality. 360p costs almost nothing and is there to test your idea and prompt. 720p suits most tasks. 1080p costs four times as much as 720p and is worth it once the scene is approved on a draft.

HDR. Gives 16-bit color that holds up to color grading. It costs twice as much, works only at 720p and 1080p, and only for five-second clips.

Loop. Stitches the end of the clip to the start so the video plays without a visible seam. Good for a website background or a screen at a trade show booth. It works only with five seconds and isn't compatible with HDR.

Duration. Ten seconds is available only without HDR and without looping.

The working routine is simple. Refine the prompt at 360p until the result starts to satisfy you, then produce the final at 720p, and switch on 1080p and HDR only for an approved scene. For reference, a five-second 720p clip without HDR costs 48 tokens, and the exact cost always shows below the input field before you launch.

One important thing to know upfront. Unlike Seedance or Wan, Ray 3.2 doesn't generate sound when working from text or a photo. Voiceover, music, and effects have to be added separately, for example with ElevenLabs and Suno, which are also in the subscription.

Where Ray 3.2 trails the competition

An honest picture matters more than hype, so let's cover this too. The independent lab Curious Refuge tested Ray 3.2 across five criteria and called it a clear step forward from the previous version 3.14, especially in cinematic quality and prompt adherence. In their assessment, though, the model doesn't yet reach the level of Seedance 2.0 and Kling 3.0.

Ray 3.2 is weakest where a scene calls for complex physics or cartoon-style animation. The further a request drifts from a realistic scene with a clear subject and clear motion, the wider the gap between the idea and the result.

The takeaway is simple. Reach for Ray 3.2 for realistic atmospheric scenes, product shots, and animating photos, especially if color grading comes next. For complex stunts, animation, and clips with sound, other models fit better.

Picture Quality Comes First

Luma Ray 3.2 doesn't try to be a model for every situation. Its strength lies in a cinematic image you can grade like real footage, careful prompt adherence, and calm, clean motion without the telltale AI jitter. There's no sound and complex physics is harder for it, but wherever the image itself matters most, it performs very well.

Ray 3.2 works best alongside other tools. Shoot the scene here, add sound in ElevenLabs or Suno, and for clips where a character needs to speak, reach for Seedance or Wan. A single model rarely covers an entire clip.

That's exactly what unitool.ai is for. All of these models sit in one subscription, no foreign cards and no separate sign-up on each platform. Start with a 360p Ray 3.2 draft, dial in the prompt, then produce the final scene at whatever quality you need.