Same Omni video twice: no seed, but a replay returns it
Omni Flash lists no temperature or seed, and Sume rejects seed on every video model. To repeat a clip, replay an idempotency key or use a video_url edit.

You cannot reproduce a Gemini Omni Flash clip by setting a seed. Google's Omni page lists no system instructions, temperature, top_p or stop sequences, and Sume's video docs say no v1 model accepts seed: each model reports seed: false and the API rejects the field. What Sume does give you is a replay: send the same Idempotency-Key with the same payload and you get the original job back, not a new generation. Read on 2026-10-08.
So there are two different needs here. If you want the identical file again, replay or re-download. If you want the same clip with one thing changed, use an edit.
Which control exists where
The table separates what a sampling control would do from what each route offers.
| Need | Google Omni docs | Sume video docs |
|---|---|---|
| Seed | Not listed; no temperature or top_p | seed: false on each model; field rejected with 400 |
| Same file again | Keep the interaction (store on) and the output | Same Idempotency-Key returns the original job |
| Change one detail | Multi-turn edit with previous_interaction_id | video_url edit; prompt carries the change |
| Pin the look | Reference images and first/last frame | image_url, end_image_url, up to 10 reference images |
Replay without paying twice
Sume's differences table lists idempotency as an addition: send Idempotency-Key to make retries safe, and a replay returns the original job. A key reused for a different payload is a 409 idempotency_conflict, which is the right failure when someone changes the prompt and forgets the key.
One practical rule: build the key from the clip, not the attempt. A key such as spring-hook-03-720p stays stable across crashes. A draft at 360p and a final at 720p need separate keys, since they are different payloads.
Edit instead of re-rolling
Re-rolling an Omni clip to fix one detail usually changes everything else too. Sume's Omni row has an edit mode: send video_url and a prompt such as Replace the bottle with an apple. Keep everything else the same. Resolution is optional and defaults to 720p, and aspect_ratio is not accepted for an edit. Google's page also recommends the phrase Keep everything else the same for targeted edits.
An edit bills output seconds at the same per-second rate. When the caller sends no duration, the reserve assumes 8 seconds: $1.00 at 720p on Sume. Google's direct list is about $0.10 per second at 720p.
What to store
Store the prompt, the resolution, the idempotency key and the job id next to every approved clip. If a client asks for the same clip next month, the file is the answer. A new request is a new roll.
Keeping a result you like
Since you cannot repeat a generation, save the file the first time. Download the result as soon as the job completes and store it with the prompt. For small changes to a clip you like, use edit mode with video_url and a prompt instead of generating again.
Sources
Related posts
More in Models
- Seedance 2.5 has a 4-second minimum: the shortest clip is $1.0747
Seedance 2.5 on Sume takes 4 to 30 seconds. What a 4-second clip costs at 480p, 720p and 1080p on each Seedance tier, and where shorter hooks go.
- sume/music-auto or pinned lyria-3.5: what changes in the job
Music Router takes sume/music-auto (Lyria 3.5 today), lyria-3.5 or lyria-3-pro. What job.model and routed_model echo back, and why the price stays the same.
- Swap four people in one video: H3 Max Recast bills per second
h3-max-recast takes one source video and 1 to 4 person photos, 5 to 30 seconds. Price is $0.375 or $0.5625 a second whether you swap one person or four.
- Hy Image 3.5 Preview is not in Sume's catalog: what to call instead
Sume's image catalog does not list Hy Image 3.5 Preview. Check with one jq call, then pick a listed model by price: Soul, gpt-image-2.5, Seedream or Flux.
Written by Sume