Recurring cast in Gemini Omni Flash 1.1: 10 images, 3 short clips

For a recurring cast in a vertical episode, Gemini Omni Flash 1.1 on Sume takes up to 10 reference images and up to 3 reference clips of 3 seconds or less in one request, and you point at them in the prompt as <IMAGE_REF_0> and <VIDEO_REF_0>. A shot runs 3 to 10 seconds in 9:16 or 16:9 with native synced audio. At 720p the billed rate is $0.125 a second, so a 10-second shot is $1.25.

What the model accepts

How to write a cast into the prompt

Number your references in the order you send them and give each a name in the prompt: 'MAYA is <IMAGE_REF_0>, JUN is <IMAGE_REF_1>.' Then write the shot with those names. Use one image per character per angle you care about, and keep the three short clips for a motion you want repeated, such as a walk or a gesture. Because clips must be 3 seconds or less, trim them first with the video-trim endpoint.

Pick the references that carry identity. A clean front view of each face does more than a busy still of the whole scene, and a prop you need on screen deserves its own image.

The three clips are for motion rather than looks. If a character has a recognisable walk, give a 3-second clip of it as <VIDEO_REF_0> and name it in the prompt.

Cost per shot

The list rates are $0.03, $0.10, $0.15 and $0.30 a second at 360p, 720p, 1080p and 4K. Sume bills the provider list times 1.25, which gives $0.125 a second at 720p and $0.1875 at 1080p. A 10-second shot at 720p is therefore 10 x $0.125 = $1.25. A 60-second episode built from six such shots is $7.50 before any retries, and the Timeline fee adds $0.10 per output minute.

These figures use the repo's Omni list rates and the 1.25 factor; a retry is a second charge.

The request

Send the references as input_references on POST /v1/videos. The model needs no generate_audio toggle because audio is always on. If you also send frame_images, Sume uses them and drops the references, so choose one approach per shot.

curl -sS -X POST "https://api.sume.com/v1/videos" \
  -H "Authorization: Bearer $SUME_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gemini-omni-flash-1.1",
    "prompt": "MAYA is <IMAGE_REF_0>. She opens a locked door and gasps.",
    "duration": 8,
    "resolution": "720p",
    "aspect_ratio": "9:16",
    "input_references": [
      { "type": "image_url", "image_url": { "url": "https://media.sume.com/maya-front.png" } }
    ]
  }'

Limits to plan for

Ten images is a ceiling for the whole shot, not for the series. A scene with three people and two locations is fine; a cast of eight with props is not. Split a crowded scene into two shots and cut them together. Also remember that the edit mode via video_url is a Video Router feature and is not part of the OpenRouter-shaped /v1/videos body.