What is an AI video seed? How to repeat a result without one

An AI video seed is a number that sets the random starting point of a generation. On a generator that accepts one, sending the same seed with the same prompt and settings is meant to reproduce a result, and a new seed gives a new variation. Sume's video models don't accept a seed: each one reports seed: false, and a request that sends seed is refused.

So on Sume you keep a result by storing the clip, keep a look by pinning the first frame, and retry safely with an idempotency key. The facts below come from the Video generation docs and the API code, read on 2026-09-27.

What does a seed do in AI video generation?

It makes randomness repeatable. A generator draws random values to start each clip, and the seed picks which ones. Keep the seed and every other input the same, and a model that supports seeds is meant to produce the same output; change only the seed, and you get a different take of the same request.

Support varies. Sume's docs list seed in the request table as "Seed for deterministic generation (not guaranteed by all providers)", then note in their list of Sume differences that no video model accepts it.

Can I set a seed on Sume's video models?

No. In GET /v1/videos/models, every model reports seed: false, and POST /v1/videos refuses the field instead of ignoring it: today a request with seed returns 400 unsupported_parameter with the message “seed is not supported in v1 — every model reports seed: false.” Provider options are no way around it either, because allowed_passthrough_parameters is empty for every model and a non-empty provider.options is refused the same way.

The fix is to remove the field. Video generation API 400 errors covers the error body and the other refused parameters.

How do I get the same result again without a seed?

Decide what you need to keep, then hold that part fixed:

Does the same prompt give the same video?

Sume's docs make no such promise, and there is no seed to hold fixed. Treat each generation as a new take and keep the ones you want.

Resending a request with the same Idempotency-Key doesn't make a second take either: the replay returns the job you already have instead of generating again. More on keys in Idempotency keys for AI video APIs.