Luma Ray 3.2 Image to Video

Luma's Ray 3.2 is one of the hosted models people actually get excited about, and this node is its image-to-video entry point: give it a start frame, optionally an end frame, and get a video back. It's a partner/video API node built into ComfyUI core - nothing to install, nothing running on your GPU. The whole generation happens on Luma's servers via Comfy's proxy, billed per render through your Comfy account.

The headline constraint is in the node description and it's non-negotiable: image-anchored generations are always 5 seconds. There's no duration slider because there's no duration choice. If you want 10 seconds, you use the text-to-video sibling or the keyframes node. That's a deliberate design decision on Luma's part - anchored motion is expensive to extend - and it means you plan shots as tight 5-second beats rather than long takes.

The inputs are refreshingly minimal. prompt describes the motion you want, resolution goes from 360p up to 1080p (default 720p, a sane default for testing). start_frame and end_frame are both optional - but at least one is required, and the node will error if you provide neither. The classic play is both: a product shot morphing into its hero angle, a character walking from pose A to pose B with identity pinned at both ends. A single start frame is the other common mode, essentially animated still photography.

The one control that does more than it looks like: loop. Flip it on and Ray 3.2 makes the video loop seamlessly - the end wraps back to the start. It's a genuinely useful effect for background plates, character idle animations, or anything destined for a website hero section. But you can't have both looping and an end_frame, because a loop by definition has no separate ending. The node will reject the combination rather than silently ignore your end frame, which is the right kind of failure.

Outputs are VIDEO plus a generation_id string. That ID is the receipt for the job - handy if you're logging renders or building a pipeline that needs to track generations, though for casual use you can ignore it. seed does the standard thing here: it only decides whether the node re-runs, because Ray 3.2 output is nondeterministic regardless.

The practical rhythm: use the cheap settings (360p/540p) to test the motion you're after, then re-run at 1080p once the composition works. The price badge updates with your resolution choice, and 1080p at 5 seconds is a real chunk of credits per iteration - don't iterate at max resolution. Also, as with every hosted node, your frames leave your machine, so keep the input images appropriate.

Ray 3.2 support landed in core in June 2026, and this node is the simplest on-ramp to it. For anything that needs a longer take, a defined middle, or more than two anchors, the keyframes-to-video node is the step up - this one is the quick, reliable draft button.