Omni reference-to-video with 3 clips of 3 s: a 10 s output is $1.25

Gemini Omni Flash 1.1 (gemini-omni-flash-1.1) accepts up to 10 reference images and up to 3 reference video clips, each at most 3 seconds, so the most reference footage one request can carry is 9 seconds. The output is 3 to 10 seconds. A 10-second 720p result costs 10 x $0.125 = $1.25 on Sume, and the reference media are not a separate charge in the docs.

Request fields

On the Video Router you send reference_image_urls and reference_video_urls. In the prompt you point at them with 0-based tokens in list order, <IMAGE_REF_0>, <IMAGE_REF_1>, <VIDEO_REF_0>, and so on. The reference media are sent before the prompt.

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-r2v-001" \
  -d '{
    "model": "gemini-omni-flash-1.1",
    "prompt": "<IMAGE_REF_0> walks into the room from <VIDEO_REF_0> and sits down",
    "reference_image_urls": ["https://example.com/person.png"],
    "reference_video_urls": ["https://example.com/room-3s.mp4"],
    "resolution": "720p",
    "duration": 10,
    "aspect_ratio": "16:9",
    "mode": "async"
  }'

Limits and price at each tier

The limits below come from the Video Router docs and the model's constraints. Price is for a 10-second output.

Pitfalls

Sending a lone reference image with no first or end frame is treated as reference-to-video. Sending a video_url is different: it selects edit mode, and you cannot combine it with image_url, end_image_url or any reference_*_urls. Native audio is always on and reference_audio_urls is not available, so a voice or soundtrack you want must come from the prompt, not an audio reference.

If a 3-second reference clip is a rough cut, trim it first. The video trim API costs $0.02 per job.

Choosing between images and clips

Use reference images when you need identity: a face, a product, a logo. Use reference clips when you need motion or a place. Each reference clip can be at most 3 seconds, so cut the best three seconds of a longer take first.

Because reference_audio_urls is not accepted, a voice sample cannot guide the audio. If a specific voice matters, generate on a row that takes audio references, such as Seedance 2.x, Wan 3.0 or MiniMax H3, whose docs list audio among their input reference types.