Luma Ray 3.2 Keyframes to Video

The two-frame bookending nodes in this family are great, but what if you want three beats? Or six? That's what this node is for. It's Luma Ray 3.2's keyframes-to-video API, built into ComfyUI core, and it takes a chain of guide images - each anchored to a position on the output timeline - and interpolates a video through all of them. It's the closest thing this category has to an animatic: you storyboard the moments you care about, and the model fills in the motion between.

It doesn't do this alone. The keyframes input expects a LUMA_RAY32_KEYFRAME chain built with the sibling Luma Ray 3.2 Keyframe node: connect an image, choose where on the timeline it applies (either as a fraction 0.0–1.0 of the video or as an absolute time in seconds), and chain several together. The chain feeds into this node. You need at least 2 keyframes - the node refuses anything fewer with a message pointing you to the image-to-video node for single-frame jobs - and you're capped at 64, which for most people is not a real ceiling.

What the source quietly tells you, and what will save you a confusing afternoon: the order you chain keyframes in doesn't matter. The node sorts them by position. So chain them in any order, give each a distinct spot on the timeline, and the model places them correctly. What it won't tolerate is two keyframes resolving to the same frame - at 5 seconds the timeline is 120 frames, at 10 seconds it's 240 (both at 24fps), so keep positions spaced out. And if you set a keyframe's time past the end of your chosen duration, it errors rather than silently dropping it. The system is strict, and the strictness is the point: it catches your mistakes instead of producing a confusing clip.

The other inputs are the standard Ray 3.2 set: prompt describing overall motion and style, resolution (360p–1080p), duration (5s or 10s), and seed as the re-run control. Outputs are VIDEO plus a generation_id receipt string.

Honest expectations: this is the most powerful and the most expensive Luma node, and the results are correspondingly hit-or-miss. Keyframe interpolation across very different images can produce strange morphs - the model has to invent transitions between them. It works best when your keyframes are visually related (same subject, evolving pose or composition) and your prompt explains the through-line. Draft the interpolation at 360p or 540p before spending credits on the 1080p version; a keyframe sequence is a real chunk of billing per run.

Ray 3.2's keyframe system landed in core in June 2026 alongside the rest of the family. If you've ever wanted to direct a video instead of just prompt one, this is the node that lets you do it with a handful of stills and a timeline.