Luma AI released Ray 3.2, the next generation of its inference-style video-generation model. The biggest change: Ray 3.2 outputs native HDR and EXR-format video directly, integrating with the existing post-production pipeline (DaVinci Resolve, Nuke, Flame) without the round-trip through standard dynamic range.
The technical details: Ray 3.2 trains a dual-track generation model — the "low-dynamic-range track" handles standard 8-bit sRGB output, and the "high-dynamic-range track" generates 16-bit EXR with full latitude. The two tracks share the underlying diffusion backbone but have independent output heads, and the HDR head is trained on a curated HDR film-stock dataset.
The "inference-style" highlight: Ray 3.2 can do explicit "inference" during generation — i.e., it can predict what the next 2-3 seconds of video will look like, and use that prediction to correct the current frame. The result is dramatically improved temporal consistency and physics realism, particularly in fast-motion scenes.
The commercial angle: Ray 3.2's API is priced at $0.08 per second of generated video, with HDR output adding a 30% premium. The first batch of customers includes Pixar, ILM, and a number of advertising agencies. Luma has also opened a "studio plan" — a flat $5,000/month for unlimited generation.
The bigger signal: video generation is moving from "cool demo" to "production tool." Ray 3.2's HDR/EXR support and inference-style correction bring it into the workflow of professional post-production, and the Luma-Pixar-ILM triangle is a clear sign that "AI video" is being adopted by the traditional film industry.