Luma AI drops Ray2 video model with 10x compute

Anyone who spent time with last year's AI video generators knows the painful reality: almost every cinematic clip looked like it was filmed in slow motion underwater.

Luma AI rolled out Ray2, the company's next-generation video foundation model powering Dream Machine. Built on a fresh multi-modal architecture, Ray2 was trained directly on raw video data using 10 times the compute power thrown at the original Ray1.

Real physics, real speed

The most noticeable change here is the death of artificial slow-mo. Previous models leaned on sluggish pacing to mask glitchy artifacts. Luma claims Ray2 eliminates those slow-playback issues, delivering fast, coherent motion where characters and objects actually interact according to physical laws.

On the spec sheet, the base clips clock in between 5 and 10 seconds, but users can chain smart extensions up to 30 seconds without losing quality or motion continuity. It supports 540p, 720p, and 1080p native exports, alongside a 4K upscaler and prompt-driven audio generation.

To help creators steer the beast, Luma introduced a standardized prompting framework spanning seven blocks: subject, action, details, scene, style, camera movement, and reinforcer. Alongside crane, dolly, and tracking shot controls, you can also trigger seamless looping directly from the UI.

Ray2 is available right now for paid Dream Machine subscribers across both the web platform and the iOS app, with an API open for developers.

Why it matters

For commercial creators and video editors, Ray2 tackles the biggest obstacle keeping generative video out of real production pipelines: consistency. Most AI video clips look great for three seconds as B-roll before melting into chaos. If Ray2 can reliably hold its coherence over 30-second clips at natural speeds, it transforms from an expensive party trick into a legitimate production asset.

Across the industry, this release proves the video compute race is getting brutal. By throwing 10x more compute at native video training, Luma is signaling that lightweight workarounds have hit a ceiling. Competitors like Runway, Pika, and Kling will have to burn serious GPU hours just to stay in the conversation.

We think Luma's focus on workflow pain points—fixing slow-motion playback, adding real camera directions, and standardizing prompt syntax—is genuinely smart. But ramping up compute tenfold comes with massive operational overhead. Dream Machine will need to prove it can keep generation queues reasonable and API costs sustainable once production studios start hammering it with 4K renders.