Day-to-night and season-change video transitions with Omni on Sume

For a day-to-night or season-change transition, photograph or generate the same scene twice, send the day version as the first frame and the night version as the last frame to Gemini Omni Flash 1.1, and let the prompt describe the time passing. An 8-second clip is $1.00 at 720p and $1.50 at 1080p on Sume.

The model can fill the time between two keyframes because Google's first and last frame control generates continuous video between them, with smooth transitions (Google blog, read 2026-10-05). On Sume you send the pair as frame_images, or as image_url and end_image_url on the Video Router.

Build the pair

A time transition works only when everything but the time stays put. Fix the camera on a tripod or render both from one fixed viewpoint. Match crop, horizon and framing. If you are generating the frames with an image model, produce the day image first and edit it to night, rather than generating two separate images, so that the geometry matches.

For a season change, keep the same trees and buildings and change foliage, snow and light. The farther apart the two frames are, the more the middle depends on the model's guess, so a spring-to-summer pair is safer than summer-to-winter.

Prompt the passage of time

Describe how the change happens: "Time-lapse, the sun sets, windows light up one by one, clouds streak across, the camera is locked off." Name what does not move. If you want the lights to turn on, say so; the last frame alone does not tell the model in what order.

• Day to night: sky gradient, window lights, street lamps, shadows lengthening.
• Summer to winter: leaves fall, snow accumulates, light cools.
• Construction or renovation: pair a before image and an after image, and name the work in the prompt.
• Keep the camera locked off unless you want a move as well.

What it costs

Clip length controls how gradual the change looks. Eight seconds is a good starting point; the Omni range is 3 to 10.

Seasons and light in one frame pair

Time is the most legible kind of change for a viewer, which makes it a forgiving first project with first and last frame control. The viewer already expects the sky to darken and the lamps to come on, so small errors in the middle read as artistic license instead of mistakes. The risk is geometry. Any object that appears in one frame and not the other, such as a parked car or a person, has to be created or removed in the middle. Clear the scene of movable things before you shoot or render the pair.

If you must have a person in the scene, keep them in the same pose in both frames, or use a reference image workflow instead.

Sound

Native audio is always on for this model, and there is no generate_audio toggle: the API rejects false. A day-to-night clip may therefore arrive with ambient sound that changes with the light. If you want a specific sound bed, such as birdsong fading into crickets, mix it yourself over the clip afterwards, or prompt it directly and check the result.

Where it fits

A pair of photos and a short prompt gives you one clip, and the draft route keeps experiments cheap: ten 360p variants of an 8-second transition are $3.00.

A cheap test plan

Run the pair at 360p for 8 seconds ($0.30). Check three things: does the camera stay locked, do objects stay in place, and does the change feel gradual rather than a dissolve. If yes, run a 1080p final for $1.50. If the middle looks like a crossfade, add more motion words to the prompt or lengthen the clip, and run another draft for $0.30.