Indie game trailer: 40 s of Omni with 3-second gameplay references

To build a 40-second indie game trailer on Sume, cut up to three 3-second gameplay clips per shot with video trim, pass them to gemini-omni-flash-1.1 as reference_video_urls, and generate four 10-second shots. Each reference clip may be at most 3 seconds, and you may send at most three clips per request. The trims cost $0.02 each according to the docs, so twelve trims add $0.24.

What Google and Sume each allow

Google's announcement says Omni 1.1 Flash accepts video references of up to 3 seconds to keep visual context and character consistency (Google, read 2026-10-05). The Sume catalog entry is stricter about shape: up to 10 reference_image_urls and up to 3 reference_video_urls, each at most 3 seconds, addressed in the prompt as <IMAGE_REF_0> and <VIDEO_REF_0> (0-based, list order) (Sume docs: Video Router, read 2026-10-05).

Steps

The reference clips condition a fresh generation. They are not an edit of your footage, and the model may not copy your art exactly. Treat the result as concept footage and mark it as such in the trailer description.

• Import your gameplay capture to your workspace (POST /v1/media-imports). Video trim reads only media.sume.com clips.
• Cut 3-second windows with POST /v1/video-trim: start plus duration: 3. Keep the default exact precision so the window starts where you asked.
• Pick the best 3 seconds per shot: a hero character moving, an ability, a menu transition.
• Write the shot prompt with tokens: "The knight from <VIDEO_REF_0> runs through the cave from <VIDEO_REF_1>, cinematic camera."
• Draft at 360p, check that the reference carried over, then finish at 720p or higher.

Cost sheet

Worked from the rates in Sume's estimator ($0.03 per second at 360p and $0.10 at 720p, times 1.25); verify against the catalog and pricing page.

Join and sound

Join the four finals on Timeline 1.0 with starts 0, 10, 20, 30. The timeline takes one audio spine, which suits a trailer: lay the music track as audio.url or as a soundtrack bed with ducking (Sume docs: Timeline 1.0, read 2026-10-05). Omni generates native audio on every job, so decide whether you want that sound in the cut before you assemble.

Keep your own capture for the logo and the store-page shot. Generated shots suit mood and transitions; a trailer that has to show real gameplay should show real gameplay, with the generated clips between.

What can go wrong

Reference clips of more than 3 seconds are rejected, and so are more than three of them. If your best gameplay moment is 7 seconds long, split it into windows and pick one per request. Check the request for the 0-based token order: the first clip in the list is <VIDEO_REF_0>, not <VIDEO_REF_1>. A prompt that names a token with no matching clip is a request error you can avoid by building the prompt and the list from the same array.

Another trap is audio. Omni takes no audio references and always generates its own native audio, so you cannot pass your game's soundtrack through it. Mix the soundtrack in at the timeline stage as the audio spine instead.