Which Omni input continues a scene: end frame, reference clip or edit?

On Sume, continue an Omni scene by sending the last frame of the previous clip as image_url for a continuation, or the last three seconds as reference_video_urls to carry the look; use video_url only when you mean to change the clip you already have. The three inputs do different jobs, and only the first two make new footage.

Google describes scene extension in Omni 1.1 Flash as analyzing up to 10 seconds of prior context and extending in 10-second steps up to 40 seconds (Google blog, read 2026-10-05). The Video Router page says Sume does not expose previous_interaction_id or extend, so a continuation on Sume is a new job whose inputs you choose.

The three inputs side by side

All three go to gemini-omni-flash-1.1, which routes by the shape of the request. Rules come from the Video Router page.

Option 1: last frame as the opener

Extract the final still of the previous clip with the Video frames API, then send it as image_url with a prompt that says what happens next. A still has position and light but not motion, so write the motion into the prompt: "the camera keeps drifting right as she opens the door". This is the most predictable route for a hard continuation, and it works for 3 to 10 seconds.

curl -X POST https://api.sume.com/v1/video-frames \
  -H "Authorization: Bearer $SUME_API_KEY" \
  -H "Content-Type: application/json" \
  -H "Idempotency-Key: last-frame-017" \
  -d '{
    "video_url": "https://media.sume.com/artifacts/artf_demo/clip1.mp4",
    "at": [9.9]
  }'

Option 2: a reference clip

Trim the last 3 seconds with the Video Trim API, at $0.02 per job, and send the result in reference_video_urls. In the prompt, address it as <VIDEO_REF_0>. Video references carry the look and character but, like any reference, do not fix the first frame of the new clip, so the cut may not be seamless. Google caps video references at up to 3 seconds as well.

Option 3: edit

video_url is for changing what you have: swap an object, change the light. It does not extend. If the clip is too short, edit is the wrong tool, because the output length follows the source.

What a 40-second scene costs this way

Four 10-second clips at 720p are $5.00, at 1080p $7.52, at 4K $15.00, plus about $0.02 per trim. The continuity depends on your prompts and on the chosen input, not on a stored context.

Common mistakes

The first mistake is sending the previous clip as video_url and expecting a longer clip. The edit route returns the same length as the source and rejects aspect_ratio, so you get a changed copy, not a continuation. The second is mixing video_url with image_url or reference fields; the API rejects the combination. The third is sending a reference clip longer than 3 seconds. Trim it first.

A fourth, quieter mistake is changing the aspect ratio between clips. Omni takes 16:9 or 9:16 only, so pick one for the whole scene and send it on every request.

Keep the audio in mind

Native audio is always on for this model, so each new clip has its own generated sound. A joined scene can have audible steps at the cuts. Plan to lay one audio spine under the joined clips on a Timeline, or accept the steps as scene changes.

Pick by the problem

• Need the exact end position to carry on: last frame as opener.
• Need the same character and style, any pose: reference clip.
• Need the same clip with one change: edit.
• Need a long unbroken take: check the length limits of other models in the Video Router before chaining 10-second clips.