Luma Ray 3.2 API: 16 keyframes, and Sume's two-frame route

Luma's Ray3.2 announcement, dated June 9, 2026, says the model takes up to 16 keyframes per clip and is available as an API. The Sume video docs read for this page name no Luma model, and the frame inputs they describe are a first frame and a last frame, so a many-keyframe sequence on Sume is a chain of two-frame clips.

What does Luma say about Ray3.2 keyframes?

Luma's announcement, read 2026-09-29, lists up to 16 keyframes per clip, 1080p, up to 20 seconds per generation, and API availability. Request fields, pricing and rate limits are not in the text saved for this page, so none are given here. Take the API shape from Luma's own documentation.

Does Sume list Luma Ray 3.2?

No. The model ids named in the Sume video docs are seedance-2, seedance-2.5, wan-3.0, minimax-h3, minimax-h3-max and gemini-omni-flash-1.1; there is no Luma entry. The authority is the live catalog: GET /v1/videos/models returns every model with its supported frame types, so check it before relying on this page.

How many keyframes can a Sume request take?

Two. frame_images entries carry a frame_type of first_frame or last_frame, and each model's supported_frame_images says which it accepts. The seedance-2 catalog example lists both.

{
  "model": "seedance-2",
  "prompt": "Camera glides from the doorway to the window",
  "frame_images": [
    {
      "type": "image_url",
      "image_url": { "url": "https://example.com/a.png" },
      "frame_type": "first_frame"
    },
    {
      "type": "image_url",
      "image_url": { "url": "https://example.com/b.png" },
      "frame_type": "last_frame"
    }
  ]
}

How do I get a longer path through many images?

Treat each pair of neighbouring images as one clip: image 1 to image 2, then image 2 to image 3. That takes one fewer clip than you have images, which is plain arithmetic rather than a Sume rule. Then join the clips with Timeline 1.0, which takes an ordered video[] list and returns one MP4 at $0.10 per output minute, rounded up.

What is lost compared with one keyframed clip?

Each clip is a separate generation job, so each clip's motion is planned on its own and only the shared image ties neighbours together. A model that takes 16 keyframes in one clip, as Luma describes, is a different shape of request. If continuity matters more than control, generate fewer, longer clips; see extending a video past 30 seconds. The video docs list per-model limits.

What does the join step need?

Timeline 1.0 wants one audio spine plus ordered video slots. The first slot must start at 0, later starts must increase, and transitions such as fade, dissolve or wipeleft go only on slots after the first, at one second or less. A render takes 1 to 200 slots, so even a 15-clip chain fits. If the chain has no soundtrack yet, set audio.mode to silence with a declared length.