Luma

Luma builds generative media models for video and image, positioned around the goal of models that can generate, understand and operate in the physical world. Two models are current: Ray 3.2 for video and Uni-1 for images, both reachable from the web app and the API.

Ray 3.2

Ray 3.2 covers four workflows in one model — text-to-video, image-to-video, Modify Video (video-to-video editing) and Reframe (aspect-ratio change).

The headline change over earlier Ray versions is multi-keyframe control. Previous releases took a single start frame, sometimes an end frame, and generated the motion between them. Ray 3.2 allows up to 16 keyframes inside a single clip, so the point where a look shifts, a movement lands or a beat hits is specified rather than inferred (lumalabs.ai/ray).

Other capabilities:

• Resolutions — 360p, 540p, 720p and 1080p
• Native 16-bit HDR generation with EXR export, aimed at post-production pipelines
• Expressive Facial Performance — up to 8 faces
• Performance Tracking — skeletal pose transfer
• Improved Motion Transfer across subjects and styles

Caution

Text-to-video and image-to-video generate no audio. Modify Video and Reframe preserve the source clip’s audio, but for the generative modes voiceover, music and effects have to be added separately.

Uni-1

Uni-1 (current release 1.1) is the image model, described as brand-aware — it generates and edits assets while holding style consistent across a set. It handles text-to-image, image editing, style transfer and multi-panel output at up to 2048 px (API docs).

API

Both models are exposed through the Luma API, with the console at platform.lumalabs.ai and reference at docs.agents.lumalabs.ai.

Model	IDs	Notes
Ray 3.2	ray-3.2, ray-3.2+hdr	up to 1080p; 20 s max for video-to-video
Uni-1.1	uni-1, uni-1-max	up to 2048 px

Ray requests are split by type: video for generation, video_edit for editing, video_reframe for aspect-ratio changes.

Dream Machine

Dream Machine was the original name of Luma’s video product and is the name most references still use. It has been folded into the main web app — lumalabs.ai/dream-machine now redirects to lumalabs.ai/app.

Open Physical AI Lab

Alongside the products, Luma runs the Open Physical AI Lab, an open-science effort aimed at generalization in physical AI, with work published openly.