Alternative ending for a Short: write your own, not a preset
YouTube's Extend with AI offers preset prompts for a new ending. With Sume you write the ending yourself and feed Gemini Omni up to three 3-second clips.

To write your own alternative ending for a Short, cut up to three clips of 3 seconds or less from it and send them to Gemini Omni Flash 1.1 as reference videos with a prompt of your own, 3 to 10 seconds long. YouTube's Extend with AI, as reported by Social Media Today, lets a creator pick a segment up to 5 seconds long and choose from suggested prompts, but not write a custom one.
The YouTube facts come from Social Media Today's report, read 2026-10-03, which is third-party. Sume's facts come from the video trim and Video Router docs.
What does Extend with AI let a creator do?
The report says the feature was announced in late September 2025 and works through the existing Cut option. A creator selects a segment up to 5 seconds long, picks from AI-generated prompt suggestions based on the content, and gets a new segment of typically 8 seconds or less that becomes a separate Short linked to the original. Music can be added from a curated list of tracks. The report lists the rollout as gradual to all countries outside the EU and UK, and states that custom endings are not possible.
- Segment in: up to 5 seconds.
- Segment out: typically 8 seconds or less.
- Prompt: chosen from suggestions, no custom text.
- Attribution: credited to the creator who generates it, with a link to the original.
How do I write my own ending on Sume?
Gemini Omni Flash 1.1 accepts reference_video_urls with up to 3 videos, each at most 3 seconds, plus up to 10 reference_image_urls. In the prompt you address them as <VIDEO_REF_0>, <VIDEO_REF_1> and so on, counting from zero in list order. A reference-video request carries no video_url, because that field is the edit mode and cannot be combined with references.
Cut the references with video trim: start plus exactly one of end or duration, so a 3-second moment is start: 12, duration: 3. Trim costs $0.02 per job and the output is a new MP4 that is never the source. Use precision: exact (the default) when the cut must land on a frame.
curl -X POST https://api.sume.com/v1/video-trim \
-H "Authorization: Bearer $SUME_API_KEY" \
-H "Content-Type: application/json" \
-H "Idempotency-Key: ending-ref-0" \
-d '{
"video_url": "https://media.sume.com/artifacts/artf_demo/short.mp4",
"start": 12,
"duration": 3
}'What should the prompt say?
Describe what happens after the last reference moment, and name each reference so the model knows which subject you mean. For example: the character from <VIDEO_REF_0> opens the box on the table and finds it empty; hold the same room and lighting. Set duration between 3 and 10, aspect_ratio 9:16, and resolution 720p, the model's default, until you like the idea.
Synced audio is always on for this model and generate_audio: false is refused, so plan on a new audio moment in the ending. If you want your original soundtrack to carry through, see splicing an AI ending onto a Short.
How do the two routes compare?
The routes differ in who writes the prompt and where the file ends up.
| Question | Extend with AI | Sume with Gemini Omni |
|---|---|---|
| Segment in | Up to 5 seconds | Up to 3 videos, each up to 3 seconds |
| Prompt | Suggested only | Your own, up to 20,000 characters |
| New clip length | Typically 8 seconds or less | 3 to 10 seconds |
| Result | A separate Short linked to the original | An MP4 on media.sume.com you can edit further |
| Price | Not stated in the report | Provider list x 1.25 per output second |
What Sume does not do
Sume does not publish to YouTube from this flow, does not add the link-back, and does not promise that the generated ending keeps a character consistent. Review the clip, and keep YouTube's own disclosure and remix rules in mind if the source footage is not yours.
Sources
Related posts
More in Use cases
- Amazon Brand Registry photos: no computer-generated mockups
Brand Registry rejects stock, digitally altered and computer-generated mockup images. Keep an untouched photo set for enrollment and use AI for listing images.
- Amazon Full-Funnel Campaigns: the video, display and audio kit
Amazon's unBoxed 2026 Full-Funnel Campaigns span video, display and audio. What a brand should have ready, and which pieces Sume can make: video, images, music.
- Prime Big Deal Days: three deal drops a day, one clip, new price text
Amazon drops new Prime Big Deal Days deals at midnight, 8 a.m. and 1 p.m. PDT. Swap price text per drop with authored caption cues, no clip re-render.
- Amazon's Prime Day video advice: product in 2 seconds, sound off
Amazon Ads' Prime Day guide says show the product in the first two seconds and design for sound-off. How to add captions and test the opening with Sume.
Written by Sume