Ray3.2 takes 16 keyframes per clip. Sume takes two frames: chain them

Luma lists up to 16 keyframes per Ray3.2 clip. Sume does not ship Ray3.2, but its first and last frame jobs chain into a keyframed shot.

5 min readSume
All posts

Luma's Ray3.2 announcement (2026-06-09, read 2026-10-05) lists up to 16 keyframes per clip, 1080p, up to 20 seconds, native HDR with 16-bit EXR export, facial performance tracking for up to 8 faces, and a Reframe tool. Sume does not ship Ray3.2; it is not in the video catalog. If the part you want is keyframe control, you can approximate it on Sume with chained first-frame and last-frame jobs and join them with Timeline 1.0. It is a different workflow with different limits, not a Ray3.2 substitute.

Vendor facts versus what Sume lists

Rows marked not listed mean the Sume docs I read do not describe the feature, not that it is impossible; check the catalog entry for the model you plan to use.

Ray3.2 (Luma) versus frame controls in the Sume docs, read 2026-10-05
ItemLuma Ray3.2Sume video routes
Keyframes per clipUp to 16First and last frame (frame_images); Omni 1.1 uses image_url plus optional end_image_url
Clip lengthUp to 20 sPer model: for example Seedance 2.5 4 to 30 s, Omni 1.1 3 to 10 s
Max resolution named1080pPer model, up to 4K on Omni 1.1
HDR / EXRNative HDR, 16-bit EXRNot listed in the Sume video docs
Face trackingUp to 8 facesNot listed

The chain: N keyframes, N-1 jobs

With key stills A, B, C, D, make three jobs: A to B, B to C and C to D. Each job gets a first frame and a last frame, and the join places them end to end. Sixteen keyframes would be fifteen jobs, so the approach suits three to five keyframes, not sixteen.

Pick a model whose supported_frame_images lists both first_frame and last_frame. The video docs' example for Seedance 2.0 lists both; read the entry for the id you pick.

curl -X POST https://api.sume.com/v1/videos \
  -H "Authorization: Bearer $SUME_API_KEY" \
  -H "Content-Type: application/json" \
  -H "Idempotency-Key: key-a-to-b-001" \
  -d '{
    "model": "seedance-2.5",
    "prompt": "Slow, smooth move from the first image to the second",
    "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" }
    ],
    "duration": 6,
    "resolution": "720p"
  }'

Planning a keyframed shot on Sume

Start by writing the keyframes as a list of stills with one line each describing the moment. Generate the stills first with an image model, then check the list as a storyboard before spending on video. Each leg then has a first and last frame that you have already approved, which is the main thing keyframe control is for.

Leg length is a choice: six seconds per leg gives a 12-second shot from three keyframes. Keep the legs similar in length for an even pace, and keep camera direction consistent across legs, or the joins will read as separate shots.

  • Approve the stills as a storyboard first.
  • Equal leg lengths give even pacing.
  • Use hard cuts between legs unless the stills differ in lighting.

Join the legs

Place the three MP4s as video[] slots with starts 0, 6 and 12. Each leg ends on the still the next leg starts from, so hard cuts should be nearly invisible. Run the unbilled POST /v1/timeline-1.0/plan before the render. The Timeline docs put render at $0.10 per output minute, rounded up.

If you need HDR delivery or face tracking across a long performance, that is a vendor-specific feature; read Luma's Ray3.2 page for current access.

Sources

Related posts

More in Comparisons

All Comparisons posts

Written by Sume