Gemini Omni with several videos: 3 references, no cross-video use

Gemini Omni Flash accepts up to 3 reference clips of up to 3 seconds each, but Google's limitations list also says that referencing or reasoning across multiple videos is not supported and may give degraded or unexpected output. Read together: give each clip its own job in the prompt, such as one character per clip, rather than asking the model to compare, merge or choose between videos.

Both lines are from Google's Gemini API: Generate and edit videos with Gemini Omni Flash, read 2026-10-02. Sume's Video Router mirrors the clip limit: reference_video_urls takes at most 3 clips of at most 3 seconds for gemini-omni-flash-1.1, and the edit source video_url is a single clip.

Are those two Google lines in conflict?

They are different things on the page. The 3 clips by 3 seconds line is an input limit on video references; the second is a warning about multi-video prompting. The page does not define the boundary between them, so treat a single clear role per reference as the safe use and test anything beyond it.

How do I use more than one reference safely?

Google's tag guide gives an example that binds one video as a character reference and one image as an object reference: The woman in <VIDEO_REF_0> is playing the violin shown in <IMAGE_REF_0>. That keeps a single video in play. On Sume the same addressing works inside the prompt, with <VIDEO_REF_0> for the first entry in reference_video_urls, counted from 0 in list order.

The example below follows the Video Router page's reference request, with one clip and one image.

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: omni-ref-001" \
  -d '{
    "model": "gemini-omni-flash-1.1",
    "prompt": "A cat inspired by <IMAGE_REF_0> walks through the setting shown in <VIDEO_REF_0>.",
    "reference_image_urls": ["https://example.com/inputs/cat.png"],
    "reference_video_urls": ["https://example.com/inputs/setting-3s.mp4"],
    "resolution": "720p",
    "duration": 8,
    "aspect_ratio": "16:9",
    "mode": "async"
  }'

What are the limits side by side?

Sume's schema allows more reference clips on other models, but names 3 for Omni.

What would I check first if a multi-clip prompt looks wrong?

Start by simplifying the input.

• Cut to one reference clip and see if the result improves.
• Check each clip is under 3 seconds.
• Make sure the prompt names every <VIDEO_REF_N> it expects the model to use.