Gemini Omni reference-to-video: 3-second clips and IMAGE_REF tags

Gemini Omni Flash 1.1 on Sume accepts up to 10 reference images and up to 3 reference videos, and each reference video must be at most 3 seconds long. You point at them in the prompt with zero-based tags in list order: <IMAGE_REF_0>, <IMAGE_REF_1>, <VIDEO_REF_0>. There is no reference audio input for this model.

These rules come from the Sume Video Router docs and catalog, as of 2026-10-08. They are the shortest reference clip limit of any model in the catalog.

A request that uses the tags

Reference media is sent before the prompt, which is why the tags count from zero in the order you list the URLs. A prompt that says <IMAGE_REF_0> walks into the room from <IMAGE_REF_1> ties each phrase to one input.

curl -X POST https://api.sume.com/v1/video-router/generate \
  -H "Authorization: Bearer $SUME_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"gemini-omni-flash-1.1","prompt":"<IMAGE_REF_0> walks into the room shown in <IMAGE_REF_1>","reference_image_urls":["https://example.com/person.png","https://example.com/room.png"],"resolution":"1080p","duration":8,"aspect_ratio":"16:9"}'

Price of a reference clip

Reference inputs do not add a line to the price. The charge is the output seconds at the chosen resolution: 8 s at 1080p is 8 x $0.15 x 1.25 = $1.50. Compare that with Seedance, where a reference video adds 15 assumed seconds to the billed length, so a short clip with a reference video costs more than the plain clip.

Common mistakes

A request with one reference image and no first or end frame is treated as reference-to-video. If you also send a first frame, you are asking for image-to-video, and the reference tags no longer describe the mode you intended. Keep the two intents in separate requests.

• Counting tags from one: the tags are zero-based, so the first image is <IMAGE_REF_0>.
• Sending a reference video longer than 3 seconds; trim it before upload.
• Sending reference audio; the model does not accept it and the catalog lists no reference_audio_urls.
• Combining video_url (edit) with reference fields; the docs say video_url cannot be sent with image_url, end_image_url or reference_*_urls.
• Setting generate_audio: false; native audio is always on and the API rejects it.

Choosing between reference models

The vendor page on fal describes the model as grounded in Gemini's real-world knowledge and physics, with camera control written in natural language (fal page, read 2026-10-08), which fits prompts that describe a camera move next to a tag.

• Need long reference footage or audio: use Wan 3.0 (5 videos of 15 s total, 5 audio clips) or MiniMax H3 (3 videos, 3 audio clips).
• Need only look and identity from stills: Omni's 10 images is the largest image count in the catalog.
• Need to change an existing clip rather than reference it: use the Omni edit mode with video_url, which cannot be combined with references.