Omni in the Gemini app takes 5 photos; the Sume row takes 10 + 3 clips

The Gemini app's Omni page says you can turn photos into a video, up to 5. The Omni row on Sume, gemini-omni-flash-1.1, accepts up to 10 reference images and up to 3 reference video clips, each 3 seconds or shorter. Those two numbers describe different surfaces, so do not plan an API batch from the app limit.

Both are read on 2026-10-07: Google's Gemini Omni page for the app, and Sume's Video Router docs for the API row.

Side by side

The app page lists features; the API docs list fields and hard limits. The rows below only compare things both sides actually state.

How references are addressed on Sume

With references, the prompt refers to each file by a token: <IMAGE_REF_0>, <VIDEO_REF_0>, counted from zero in list order. One reference image with no first or end frame is treated as reference-to-video. Sending video_url is a different mode: it is the edit source, and it cannot be combined with image_url, end_image_url or reference_*_urls.

Here is a request body in the Video Router shape that uses two image references. It is a sketch of the documented fields, not a recorded run.

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-refs-001" \
  -d '{
    "model": "gemini-omni-flash-1.1",
    "prompt": "<IMAGE_REF_0> walks past <IMAGE_REF_1> in warm light",
    "reference_image_urls": ["https://example.com/a.png", "https://example.com/b.png"],
    "resolution": "720p",
    "duration": 8,
    "aspect_ratio": "16:9",
    "mode": "async"
  }'

What a 10-reference job costs

Sume bills Omni per output second by resolution, so the number of references does not change the Omni rate in the docs: 8 seconds at 720p is $1.00 ($0.125 per second). Check the price you will pay in the live catalog, since list prices are subject to change.

• Pick the 5 most distinct photos first if you also want the app to reproduce the idea.
• Keep reference clips at 3 seconds or less; longer clips are outside the documented limit.
• Do not send an audio file as a reference to Omni; use a model whose catalog row lists audio references if you need one.

When to use the app and when the API

Use the app for one-off exploration; Google states a subscription is required there. Use the API row when you need many variants, fixed ids and a per-second price that you can budget before you submit.

Planning a batch from the API limits

The app limit matters for a person typing in a chat box. For a script, the limit that matters is the row's. Suppose you have 24 product photos and want clips that mix them. With a 5-photo ceiling you would need five or more groups. With a 10-image ceiling on the API row you can send two groups of up to 10 and have four photos left over, or three groups of 8. Either way, plan around the row's documented number and the <IMAGE_REF_n> addressing, which lets the prompt say which photo does what.

• Group photos by what they show (product, packaging, hands) before you submit.
• Number them in list order and write the prompt with the matching tokens.
• Test one group at 360p, where Omni costs $0.0375 per second, before running the rest at 1080p.

What the app page does not say

The page does not list a reference-clip limit, an audio-reference rule or a resolution, so none of those should be assumed from the API. It also marks several capabilities as new, namely scene extensions, video to video editing, multi-turn editing and avatar. Of those, Sume's Omni row documents video editing through video_url. Check the docs for any other feature before you promise it to a client.