Luma Ray 3.2 Text to Video

This is the prompt-only half of Luma Ray 3.2's ComfyUI story, and it's the one that gets the long takes. Where the image-to-video node is locked to 5 seconds, the text-to-video node gives you a real duration choice - 5s or 10s - and none of the anchoring constraints. Pure text in, video out, with the model inventing everything from your words. It's a partner/video API node built into ComfyUI core: no install, and it renders on Luma's servers through Comfy's proxy for a per-render fee.

The input surface is about as close to zero-friction as this family gets. prompt (which Luma rewards for being descriptive - this is a model that likes motion verbs and camera language), aspect_ratio across six options from 21:9 to 9:16, resolution from 360p to 1080p, and duration. That's most of it. seed is the standard re-run control on a non-deterministic model.

The control that's easy to sleep on is loop, and it's the reason a lot of people pick this node. Flip it on and the clip loops seamlessly - end meets start with no seam. It's the exact tool for background plates, looping ambience, or anything meant to sit on a landing page. The catch is printed right on the widget: looping only works with the 5s duration. Choose 10s and the loop option is quietly unavailable; the node errors if you try to combine them, so pick your trade - long single take or short seamless loop, not both.

The quality-versus-length tradeoff is worth internalizing: 10-second generations are where Ray 3.2 either shines or shows its seams depending on how much motion you asked for. Long, complex prompts over 10 seconds tend to drift; short, contained motion holds. If your idea is a slow dolly across a scene, 10s is fine. If it's a fight scene with three characters, do it in 5 and stitch.

Output is VIDEO plus a generation_id string - the job's receipt, handy for logging, ignorable for casual use. And the cost logic is the same as its sibling: the price badge follows your resolution and duration, and 1080p at 10 seconds is several times the cost of a 540p test. The sensible loop is draft at 540p, confirm the motion, then spend the credits on the 1080p final.

Two general partner-node reminders. First, everything you type is sent to Luma's servers - prompts, not just media, leave your machine, so keep that in mind for sensitive subjects. Second, there's a content filter on the hosted side that can reject requests; when that happens it fails with an error rather than a graceful empty result. Ray 3.2 hit core in June 2026, and this is the node you reach for when you want a self-contained clip with no input image and no other node in the way.