Omni extension reads 10 s of context; a Sume chain passes a frame

What Google's scene extension carries into the next 10 seconds, and what Sume carriers hand over instead: a last frame, 3-second clips or image refs.

5 min readSume
All posts

Google's Gemini Omni 1.1 Flash scene extension analyzes up to 10 seconds of the previous clip and extends in 10-second steps up to 40 seconds (Google, read 2026-10-05). Sume's gemini-omni-flash-1.1 has no extend mode. A chain on Sume hands the next clip one frame, up to three 3-second clips, or up to ten reference images, so it carries less of the prior motion than a 10-second context does.

The four carriers on Sume

The Video Router docs describe one catalog id that Sume routes by the shape of the request (Sume docs: Video Router, read 2026-10-05). Each carrier is a different request shape.

What each Sume request shape can carry from an earlier clip (read 2026-10-05)
CarrierFieldLimitWhat it carries
Last frame as first frameimage_url1 imagePose, framing, light at the cut
Last and first framesimage_url + end_image_url2 imagesWhere the new clip starts and ends
Clip referencesreference_video_urlsup to 3 clips, each up to 3 sMotion and look from short clips; tokens <VIDEO_REF_0> and so on
Image referencesreference_image_urlsup to 10 imagesCharacter and set look; tokens <IMAGE_REF_0> and so on

What is missing

None of these is 10 seconds of the preceding clip. The longest motion you can pass is 3 seconds per reference clip, with up to three clips. And the docs list the request shapes as separate capabilities: first-frame, reference and edit. A video edit with video_url cannot be sent with any of them. Pick the carrier first, then write the prompt for it.

Google's page also puts extension in the Gemini app for Plus, Pro and Ultra subscribers; this post does not cover app availability. For API work on Sume, the question is what the next request can see.

Which carrier for which seam

Get the last frame with video frames at a time just below the clip length; a request at the exact end fails with frame_time_out_of_range.

  • Same location, same action, one continuous camera move: last frame as image_url, plus a prompt that names the continuing move.
  • A character that must stay recognizable across 40 seconds: reference_image_urls with that character, in every request.
  • A movement style you already have on film: trim 3 seconds with video trim ($0.02 per job per the docs) and send it as reference_video_urls.
  • A planned transition between two looks: image_url plus end_image_url.

Budget the seams

A 40-second piece is four 10-second jobs. Each seam is a place where the carrier choice shows. Draft the seam shots at 360p first and look at seam to seam, not clip by clip: the first and last second of each draft matter more than the middle. Join the approved clips on a timeline, with a short fade only where a cut looks abrupt.

If the continuity of motion across 40 seconds is the requirement and not a nicety, say so before you build on this route. The catalog has seedance-2.5 at 4 to 30 seconds and wan-3.0 at 2 to 30 seconds in one job (Sume docs: Video generation, read 2026-10-05), which removes seams up to 30 seconds. They are different models with their own look.

Sources

Related posts

More in Comparisons

All Comparisons posts

Written by Sume