Gemini Omni reference clips: 3 videos, 3 seconds each, VIDEO_REF_0
On Sume, Gemini Omni Flash reference-to-video takes up to 10 images and 3 clips of at most 3 s each, addressed as IMAGE_REF_0 and VIDEO_REF_0 in the prompt.

Reference-to-video on Gemini Omni Flash 1.1 is how you ask for a new clip that borrows a look, a person or a motion from material you already have. The limits are tight and specific, and the prompt needs to name each reference by position.
The limits
Sume documents the 3-second limit and its request schema repeats it. Trim longer clips before you send them.
| Input | Limit | Field |
|---|---|---|
| Reference images | up to 10 | reference_image_urls |
| Reference clips | up to 3, each 3 seconds or less | reference_video_urls |
| Output length | 3 to 10 seconds | duration |
| Aspect ratio | 16:9 or 9:16 | aspect_ratio |
Address references in the prompt
References are 0-based and numbered in list order within each kind. Write <IMAGE_REF_0> for the first image and <VIDEO_REF_0> for the first clip. A single reference image with no frame field is reference-to-video, not image-to-video; use image_url when you mean a first frame.
{
"model": "gemini-omni-flash-1.1",
"prompt": "The person from <IMAGE_REF_0> walks through the door with the camera move of <VIDEO_REF_0>",
"reference_image_urls": ["https://example.com/person.png"],
"reference_video_urls": ["https://example.com/move-3s.mp4"],
"duration": 8,
"resolution": "720p",
"aspect_ratio": "9:16"
}
What it costs
The docs bill Omni by output second at list times 1.25 as a function of resolution, and describe no separate reference charge. An 8-second 720p clip is $1.00; a 10-second one is $1.25. Audio is always on and cannot be switched off, and there is no reference_audio_urls field on this model.
Sources
Related posts
More in Developers
- Gemini Omni Flash edit on Sume: the 8-second reserve hint
A Gemini Omni edit has no duration input, so Sume reserves against a hint: 8 seconds by default, up to 30 if you send one. At 720p, that is $1.00 held.
- Gemini Omni Interactions API versus Sume /v1/videos: field mapping
Google returns Omni video as base64 in an Interactions response. Sume returns 202 with a polling_url and a content URL. See the mapping and a bytes helper.
- Gemini Omni prompt template: five parts in a Python builder
Google's Omni guide names five prompt parts: framing, style, lighting, location, action. Build them in Python and send the prompt to Sume's /v1/videos.
- generate-video 409s: preview_not_ready vs preview_image_not_ready
generate-video on an avatar preview can return two 409s: the job is unfinished, or a public preview image is missing. How to tell them apart.
Written by Sume