Video edit prompt: say what stays, then what changes (Omni Flash)
A prompt pattern for Gemini Omni Flash 1.1 video edit on Sume: one change per pass, an explicit keep clause, and what the video_url route fixes for you.

For a video edit, write the change in one sentence and then add a keep clause: "Replace the bottle with an apple. Keep everything else the same." That exact pattern is the example request in the Sume Video Router docs for Gemini Omni Flash 1.1, which edits an existing clip when you send it as video_url. One change per pass and an explicit statement of what must not move is the simplest prompt that works as an edit, instead of a new generation.
This page covers how to phrase an edit, and which parts of the output you do not need to prompt for because the route already fixes them.
What the edit route fixes for you
Edit mode is an input shape, not a separate model id. Sending video_url to gemini-omni-flash-1.1 selects the edit queue, and the source clip sets the output. Sume does not send an aspect_ratio or a duration to the provider in this mode, and an aspect_ratio on an edit request is a 400. That leaves the prompt with one job.
| Item | Behavior |
|---|---|
| Trigger | video_url (the source clip) |
| Prompt | What to change, plus what to keep |
| Aspect ratio | Not accepted (400 if sent); source framing is kept |
| Duration | Source sets it; duration is only a reserve hint, default 8 s |
| Audio | Native synced audio is always on; generate_audio: false is a 400 |
| Default resolution | 720p |
| Combine with | Not with image_url, end_image_url or reference_*_urls |
A prompt in two clauses
Name the target, say what replaces it, then say what stays. The keep clause matters because an edit model can otherwise treat the rest of the frame as open to revision. Examples that follow the same pattern:
- Change the jacket from black to olive green. Keep the person, the motion and the background the same.
- Replace the logo on the mug with a plain white mug. Keep the lighting and the camera movement the same.
- Make the wall behind her light grey. Keep her clothing, her movement and the audio the same.
What a good edit prompt avoids
Avoid vague requests such as "make it better" or "improve the lighting". An edit prompt is easiest to check when you can look at the result and say yes or no. Avoid changes that depend on information the clip does not contain, such as a brand name that must appear in readable text. Avoid asking for a new camera move, because the source clip fixes the framing and timing, and the edit route does not take an aspect ratio. If a pass comes back wrong, do not pile corrections onto the same prompt. Re-run the original sentence once, since the output differs from run to run, and only then tighten the wording.
One change per pass
Resist the long list. If you need three changes, run three edits in sequence and look at each result before you stack the next. Each pass is a billed job, so a pass that drifted costs less to catch early than at the end of a chain. The trade-off against a re-roll is in editing a finished Omni clip with video_url
A reference image is not part of this mode. If you want the new object to look like a specific product photo, the edit route cannot take it, and the pairing of video_url with image fields is a 400. In that case a new generation with references is the route, not an edit.
Request and cost
At the 720p default, Sume bills the provider list of $0.10 per second plus the 1.25 house margin, which is $0.125 per second. Because the edit reserve defaults to an 8 second hint, expect a reserve near $1.00 and send a duration that matches your clip if it is shorter. For a 5 second source, that hint gives 5 × $0.125, or $0.625 before rounding to cents.
The request body is below, which is the Sume doc example with a keep clause already in the prompt.
{
"model": "gemini-omni-flash-1.1",
"prompt": "Replace the bottle with an apple. Keep everything else the same.",
"video_url": "https://example.com/clip.mp4",
"resolution": "720p"
}Sources
Related posts
More in Models
- Which Sume image models are text-to-image only, with no photo edits
Five Sume image models reject reference images: Higgsfield Soul, Imagen 4 Fast, Imagen 4 Ultra, Recraft V4 and Qwen Image Max. Prices, ratios and edit options.
- An OpenRouter-compatible video API: sume/auto or a pinned model
Sume's POST /v1/videos follows OpenRouter's video generation API field for field. Let sume/auto pick the model, or pin a catalog id like seedance-2.5.
- Image generation API with reference images: POST /v1/images
Send a prompt plus public HTTPS reference images to Sume's POST /v1/images. Pin a catalog model or send sume/auto; the catalog lists each model's limits.
- Video 1.0 and Image 1.0 are retiring soon: move to sume/auto
Sume Video 1.0 and Image 1.0 are retiring soon and already run as aliases for the Auto path. New integrations call /v1/videos or /v1/images with sume/auto.
Written by Sume