Gemini Omni 1.1 Flash: 40-Second Scenes, First/Last Frame and 4K

Short answer: Omni 1.1 Flash is Google's control update. Write the shot as a plain, timed description, mark your images with the tags the model reads (<FIRST_FRAME>, <LAST_FRAME>, <IMAGE_REF_1>), put events and lines on timecodes such as [0-3s], draft at 360p, extend in 10-second steps up to 40 seconds, then upscale the approved take to 1080p or 4K. There are no negative prompts and no system instructions - everything you want or forbid goes into the prompt itself.

What is new in Omni 1.1 Flash

• Scene extension to 40 seconds. The model now reads the last 10 seconds of the existing video as context (earlier models looked at the final second only) and adds 10-second increments, so motion and story carry across the cut.
• First and last frame control. Two images, two tags, and the model generates the continuous video between them - the cleanest way to land a camera move or a transition exactly where you planned it.
• Video references. Up to three clips of up to three seconds each guide motion or look; their audio is ignored.
• 360p drafts and 4K upscaling. Drafts render up to 60% faster and at a third of the cost of 720p; finals are 720p by default, 1080p or 4K upscaled.

The prompt structure

Omni wants natural, literal instructions, not keyword lists. The same 2026 director's brief that works for Seedance 2.5 and Veo 3.1 works here, with two Omni-specific habits: tags for media and timecodes for time.

1. Tags first. <FIRST_FRAME> and <LAST_FRAME> for a planned transition; <IMAGE_REF_N> with a role ("controls identity only"); <VIDEO_REF_N> for a motion or look you want copied.
2. One sentence of intent - what the shot is, how long, in what space and light - followed by "keep everything else the same" when you edit or extend.
3. Timecodes for every beat: [0-3s], [3-6s], one visible action per beat, and where the motion settles.
4. Camera as one explicit move and lens.
5. Audio described explicitly - room tone, effects, and dialogue in quotes with the second and the language.
6. Constraints in plain words ("no on-screen text, no extra people") - negative prompts are not supported, so exclusions live in the prompt.

A complete example (8 s, first/last frame, dialogue)

<FIRST_FRAME> the model stands at the far end of a sunlit loft, back to camera, 16:9.
<LAST_FRAME> the same model in the same loft, now facing the camera at arm's length, soft smile.
<IMAGE_REF_1> controls her identity, face and body only - not her pose, clothing, background or lighting.

Generate an 8-second continuous shot between the first and last frame in the same loft and the same morning light. Keep everything else the same: the cream linen suit, the hair, the window light from the left.
[0-3s] she turns and walks toward the camera at an easy pace, heels on the wooden floor.
[3-6s] she slows, adjusts the lapel with her right hand, eyes on the lens.
[6-8s] she stops at arm's length and settles into the soft smile of the last frame.
Camera: slow push-in at eye level, 35 mm, no zoom, no cuts.
Audio: her heels on wood, quiet room tone, a distant tram outside; at 6 s she says "Ready when you are." in English, calm; no music.
Constraints: no on-screen text, no extra people, no watermark, no morphing.

Draft, extend, upscale

1. Draft at 360p. Check timing, the settle at the end and the dialogue placement while it is cheap and fast.
2. Extend in 10-second steps. Each extension reads the last 10 seconds, so write the next beat as a continuation ("she keeps walking, the camera keeps pushing in") and repeat the identity role and the continuity line verbatim. Stop at 40 seconds per scene.
3. Upscale the approved take to 1080p or 4K. Upscaling sharpens; it does not fix a bad cut, so approve the motion at draft resolution first.

Editing existing footage follows the same rule: one plain instruction ("replace the red jacket with a cream linen suit, keep everything else the same") beats a paragraph of adjectives.

Omni 1.1 Flash versus the others

• Choose Omni 1.1 Flash to extend a scene past 8 seconds inside Google's stack, to control the first and last frame, or to edit footage with a sentence.
• Choose Veo 3.1 for 8-second photoreal hero clips with the most realistic light and up to three ingredient images.
• Choose Seedance 2.5 for 30-second one-take stories with up to 50 references and explicit roles - see the Seedance 2.5 prompt guide.
• Choose FLUX 3 Video for multilingual speech with lip sync in 5-20 second clips - see the FLUX 3 Video guide.

The full roster and prices are in AI video generators in 2026.

Common mistakes

• Negative prompts or system instructions - unsupported; write constraints in the prompt.
• Uploading an audio file as a reference - unsupported; describe the sound instead.
• Adding new dialogue while extending a clip in which someone already speaks - not possible; plan lines per scene.
• Video references longer than 3 seconds or more than three clips - trimmed or rejected.
• Extending without repeating the identity role and continuity line - the character drifts at the cut.
• Approving at 4K - draft at 360p first; upscaling does not repair motion.

FAQ

Want the tags, timecodes, camera, audio and constraints written for you? GoldenPrompts builds Omni, Veo, Seedance and Kling briefs from a few clicks. Free to start: 24 hours of everything, no card.