Launch teaser from product shots: Gemini Omni Flash reference images
Feed up to 10 product references to gemini-omni-flash-1.1 on Sume, address them as IMAGE_REF_0 in the prompt, and get a 3-10 second teaser with native audio.

For a launch teaser built from your own product photos, use gemini-omni-flash-1.1 on Sume's Video Router in reference-to-video mode: send reference_image_urls (up to 10) and refer to them in the prompt as <IMAGE_REF_0>, <IMAGE_REF_1> and so on. Clips run 3 to 10 seconds, in 16:9 or 9:16, with native synced audio always on.
How does Sume pick the mode?
Native audio is always on: generate_audio: false is rejected. There is no reference_audio_urls, so if the teaser needs a specific voice or track, add it afterwards in Timeline 1.0.
| Capability | Send | Notes |
|---|---|---|
| text_to_video | prompt | 3 to 10 s, 360p to 4K, 16:9 or 9:16 |
| image_to_video | image_url, optional end_image_url | Same envelope |
| reference_to_video | reference_image_urls (up to 10) and/or reference_video_urls (up to 3, each up to 3 s) | Address as <IMAGE_REF_0>, <VIDEO_REF_0>, 0-based in list order |
| video_to_video | video_url | Prompt describes the edit; no aspect_ratio or duration |
What does the request look like?
Put the logo and the hero product shot in the list, then say which one does what. The reference rules: video_url is an edit source and cannot be combined with image_url, end_image_url or reference_*_urls.
curl -X POST https://api.sume.com/v1/video-router/generate \
-H "Authorization: Bearer $SUME_API_KEY" \
-H "Content-Type: application/json" \
-H "Idempotency-Key: teaser-refs-001" \
-d '{
"model": "gemini-omni-flash-1.1",
"prompt": "Slow push-in on the bottle from <IMAGE_REF_0> on a dark counter, the logo from <IMAGE_REF_1> glints in the last second.",
"reference_image_urls": [
"https://example.com/bottle.png",
"https://example.com/logo.png"
],
"resolution": "720p",
"duration": 8,
"aspect_ratio": "9:16",
"mode": "async"
}'How does a teaser fit into a campaign?
Eight seconds is a cold open, not a whole teaser. Generate two or three shots with different prompts, then sequence them in Timeline 1.0 with fades and a soundtrack. Draft at 360p while you iterate on the prompt, then regenerate the keeper at the resolution you publish.
What does it cost?
Billing is the provider list price times 1.25, charged per output second by resolution. The exact figure for each resolution comes from the catalog, so read GET /v1/video-router/models/gemini-omni-flash-1.1 rather than a number in a blog post.
How should you write the prompt around references?
Name each reference by role the first time you use it, then keep to the same placeholder. If the bottle is <IMAGE_REF_0> in one sentence, do not call it the product elsewhere. References are 0-based and follow list order, so reordering the list silently changes which picture each placeholder means: reorder the prompt with it.
Keep the scene simple, since an 8-second clip holds one camera move and one beat. Shots that need a lot of text should be added as overlays afterwards instead of asked of the model; for a plate, see the compose overlay guide in the related posts.
What should you check before publishing?
Watch each shot at full size and read every label. Check that the product shape is right, the logo is spelled correctly, and the sound fits the mood, since native audio is generated with the picture and cannot be switched off. If the audio is wrong, replace it in Timeline 1.0 by giving the render your own spine.
Keep the references clean: a plain background on the product shot and a flat logo file tend to give the model less to misread, though results are not guaranteed. Names and prices belong in your own overlay or captions, where you control the exact text.
What does this not do?
References guide the picture; they do not guarantee a pixel-exact bottle or a legible logo. Check the label and any text on the product in every shot before publishing. If you need an exact packshot to stay fixed, use the packshot as a first frame instead and read the first-frame guides in the related posts.
Sources
Related posts
More in Use cases
- Product photo on a white background: one image edit call on Sume
Turn a messy product photo into a clean white-background shot with one POST /v1/images edit. Which model, which fields, and what to check before you publish.
- Re-roll one shot in a stitched film without redoing the rest
Regenerate a single bad shot, swap its source_url in the Timeline 1.0 video array, and re-render the cut for $0.10 per output minute. The other shots stay.
- Listing photos to a narrated tour video with TTS and Timeline
Turn ten listing photos into a 45-second narrated tour: write the script, get word timestamps from TTS, then render stills on a Timeline spine for about $0.15.
- Recast a video longer than 30 seconds: trim, recast, rejoin
H3 Max Recast takes 5 to 30 seconds of source. For a 90 second video on Sume, probe it, trim ranges, recast each at 768p, then plan the join. Cost math inside.
Written by Sume