AI video seed on Sume: no model accepts one, lock shots with frames

No Sume v1 video model accepts a seed: each model reports seed: false in GET /v1/videos/models, and the field is rejected. To get a repeatable look anyway, pin the things the API does let you pin: the model id, the first and last frames, reference images, resolution, aspect ratio, duration and a fixed prompt.

That is a control strategy, not determinism. The docs say even where a seed exists, the result is not deterministic on all providers.

What the docs say

The Video generation docs list seed as an optional integer for deterministic generation, then add that results are not deterministic on every provider. In the Sume differences table: no v1 model accepts seed, and each model reports seed: false and rejects the field. Likewise provider.options must be empty, and size returns 400 unsupported_parameter.

Replace the seed with anchors

Do not use Auto if you need to repeat

model: sume/auto lets Sume choose the family. The docs say resolution is a pure function of the normalized request and the catalog version, so an idempotent replay gets the same price and the same route, but a later catalog version can route differently. If you need the same engine each time, pin a catalog id.

Pinning matters when you make a sequence: if shot 1 ran on one id and shot 2 on another, the look shifts even with the same prompt.

Idempotency is not reproduction

An Idempotency-Key makes a retry safe. A replay with the same key returns the original job, so you get the same clip again. A new key submits a new job and a new sample. So keep a table of key to clip, and when a client says shot 7 was the good one, use its output rather than trying to regenerate it.

{
  "model": "seedance-2",
  "prompt": "Product on a marble counter, soft morning light, slow orbit",
  "duration": 6,
  "resolution": "1080p",
  "frame_images": [
    {"type": "image_url", "image_url": {"url": "https://example.com/hero.png"}, "frame_type": "first_frame"}
  ]
}

A note on Ray 3.2

Luma lists up to 16 keyframes per Ray 3.2 clip (read 2026-10-05) as its control surface. Sume does not run that model. The closest Sume equivalent is the pair of frame anchors above, with a chain for more. Freeze the approved outputs as files in your workspace, since regenerating is a new sample.

How to approve a shot

Treat generation as casting: run two or three takes with different keys, pick one, and freeze it. Store the job id, the clip file and its inspect probe. Later edits happen on that file with trim, filter and Timeline, none of which re-sample the model. Only regenerate when a client rejects the shot, and then expect a different clip.

If a model's first frame is the same still and the prompt is the same, the composition starts the same, but motion is free to differ.

Quick rules

• Do not send seed; it is rejected on every v1 model.
• Pin the model id; do not use sume/auto when you need the same engine.
• Anchor with first_frame (and last_frame where listed).
• Store the approved job id and file; do not regenerate it.
• Use a new Idempotency-Key only when you want a new sample.

Before you build

Before you build, read the linked Sume docs page for the exact request fields, limits and prices, because those pages are the source of truth and can change. Run one short, cheap test with your own material first, check the output in a player and in your editor, and only then scale to the full shot list. Keep every job id and file you approve, so a later change never forces you to regenerate work that was already signed off. Note that this post describes Sume's catalog and tools; Sume does not run Luma Ray 3.2, and nothing here claims HDR or EXR output.