The best Gemini Omni Flash prompts read like a short shot description: who or what is on screen, what happens, how the camera frames and moves, what the light and style feel like, and what you should hear. Add "single continuous shot" when you want one take, name the audio you want (or say "no music"), and keep follow-up edit prompts short with "keep everything else the same." Below are the tips Google's own docs recommend plus six copy-ready example prompts.

How Gemini Omni Flash reads a prompt

Gemini Omni Flash is Google's fast video model from the Gemini Omni family, announced at Google I/O 2026. It takes any mix of text, images, audio and video as input and returns a video with sound. According to the Gemini API documentation, each generation runs 3 to 10 seconds at 24 fps, in 16:9 (the default) or 9:16, with resolutions from a 360p draft up to 1080p and 4K on the current 1.1 release.

Three behaviors shape how you should prompt it:

  • It reasons about your scene. Google's official prompt guide says you don't have to be as prescriptive as with older video models. State the intent and the important details; the model fills in plausible context from its world knowledge.
  • It makes sound by default. Speech, sound effects and ambience come with every clip. If you don't describe the audio, the model chooses it for you.
  • It can cut between shots. The API docs recommend phrases such as "in a single continuous shot" or "no scene cuts" when you want one unbroken take.

The anatomy of a strong Omni Flash prompt

Google's prompt guide groups its advice into a handful of building blocks. Use them as a checklist, not a template you must fill completely.

ElementWhat to writeExample phrasing
Subject and actionWho or what is on screen and what happens, in order"a barista pours steamed milk into a latte, forming a leaf pattern"
FramingWide, medium or close-up"tight close-up at cup level"
Camera motionHow the camera moves, or that it doesn't"slow push in", "locked-off static shot", "one continuous shot"
StyleThe feel of the footage"grounded, documentary realism" or "warm claymation"
LightingLight source and quality"soft window light from the left, warm morning tone"
LocationThe setting in a phrase, not a paragraph"a small neighborhood cafe at opening time"
AudioMusic, ambience, dialogue or silence"quiet cafe ambience, gentle milk hiss, no music"
TimingOptional beats with timecodes"[0-4s] ... [4-8s] ..."
A cinematographer's desk with a storyboard, a camera monitor and colored lighting gels arranged for planning a short video shot
A cinematographer's desk with a storyboard, a camera monitor and colored lighting gels arranged for planning a short video shot

A practical order is subject and action first, then framing and camera, then style and light, then audio. Front-loading the action keeps the model focused on what matters most in a 10-second clip.

Six example prompts you can copy

Each prompt below is written for a single 10-second generation. Swap the subject for your own and keep the structure.

1. Single-shot product hero

AI Prompt RecipeFlux.1 Pro
16:9
“A matte black ceramic coffee mug rotates slowly on a wet slate surface. Tight close-up, locked-off camera, in a single continuous shot with no scene cuts. Moody studio lighting with one soft key light from the left and a thin rim light behind. Steam curls up from the mug. Audio: quiet room tone and a soft ceramic clink at the start, no music, no dialogue.”

Why it works: it names the shot type, forbids cuts, and controls the soundtrack, so you get a clean loop-friendly product shot instead of a mini montage.

2. Timed three-beat sequence

AI Prompt RecipeFlux.1 Pro
16:9
“A trail runner on a misty mountain ridge at sunrise, cinematic and grounded. [0-3s] Wide shot: the runner crests the ridge, breath visible in the cold air. [3-6s] Medium tracking shot alongside the runner as the sun breaks through the clouds. [6-10s] The runner stops at the edge and looks out over a valley of fog. Audio: wind, footsteps on gravel, a slow building ambient synth pad.”

Why it works: the API docs support natural-language or timecode timing. Timecodes keep each beat from bleeding into the next.

3. Animate a reference image

Upload one image (a product shot, an illustration or a character design), then prompt the motion rather than the appearance:

AI Prompt RecipeFlux.1 Pro
16:9
“Turn this image into realistic footage. Keep the character's outfit, colors and proportions exactly as shown. She walks toward the camera through a rainy neon-lit street at night, medium shot, slow dolly back. Reflections shimmer on the wet pavement. Audio: light rain, distant traffic, no music.”

Why it works: the reference carries the look, so the text spends its words on motion, camera and sound. Google's guide notes that references help keep characters and objects consistent.

4. Style transformation of existing motion

AI Prompt RecipeFlux.1 Pro
16:9
“Restyle this video as handmade claymation with visible fingerprint texture on every surface. Keep the original camera movement, timing and character actions the same. Warm, soft lighting. Keep the existing audio.”

Why it works: the prompt guide describes reframing existing motion in a new look (anime, claymation, watercolor) while preserving the original details. Saying what must stay the same is as important as saying what changes.

5. Audio-led mood piece

AI Prompt RecipeFlux.1 Pro
16:9
“A lighthouse on a rocky island during a storm at dusk, wide static shot, waves crashing against the rocks, the beam sweeping through heavy rain. Painterly cinematic realism, cold blue tones with a warm beam. Audio: roaring wind and surf, a low tinny radio broadcast playing faintly from inside the lighthouse, no dialogue.”

Why it works: Google's docs encourage naming the audio as specifically as the visuals, down to the texture of a sound. A specific description gets you a specific soundtrack.

6. Seamless loop from first and last frames

On the 1.1 release you can supply a first and last frame and let the model fill the motion between them. Use the same image for both to get a loop.

AI Prompt RecipeFlux.1 Pro
16:9
“Use the first and last frame as the start and end of a seamless loop. A paper boat drifts in a slow circle on a still pond while cherry blossom petals fall around it. Static overhead camera, soft overcast light. Audio: gentle water lapping and birdsong, no music.”

Why it works: Google highlights first/last frame control for smooth transitions, camera moves and looping. A static camera makes the loop point invisible.

Six glowing video frames arranged on a dark editing table, each showing a different scene: a coffee mug, a mountain runner, a rainy neon street, a claymation figure, a lighthouse in a storm and a paper boat on a pond
Six glowing video frames arranged on a dark editing table, each showing a different scene: a coffee mug, a mountain runner, a rainy neon street, a claymation figure, a lighthouse in a storm and a paper boat on a pond

Try your prompts in the Video Studio

How to write edit prompts that don't break the clip

Conversational editing is the headline feature. In the Gemini app you edit in the same chat; through the API you pass the previous interaction's ID. The rule from Google's docs is simple: keep edits short. Overly descriptive edit prompts cause unintended changes elsewhere in the video.

Good edit prompts are one sentence of change plus one sentence of protection:

  • "Add a small dog that runs in and sits at her feet. Keep everything else the same."
  • "Change the background to a snowy forest. Keep the subject, camera and audio the same."
  • "Make it golden hour instead of midday. Don't change the action."

Habits that help:

  • One change per turn. Stack edits across turns instead of packing three changes into one.
  • Name what stays. "Keep everything else the same" or "keep the camera and timing" protects the rest of the shot.
  • Use references for swaps. To replace a character or object, attach an image of the replacement instead of describing it at length.
  • Know the limits. Google's model card lists consistency across edits as an ongoing challenge, and the API docs list voice editing as unsupported. If the dialogue needs to change, regenerate.
A film editor's dark workstation with two monitors showing the same street scene before and after a background change, lit by soft amber desk light
A film editor's dark workstation with two monitors showing the same street scene before and after a background change, lit by soft amber desk light

Controlling sound, dialogue and on-screen text

  • Music: describe it plainly ("upbeat electronic beat", "calm solo piano"). If you want silence, say "no music."
  • Removals: the docs suggest simple negatives like "no dialogue" or "no extra sound effects."
  • Dialogue: put spoken lines in quotes and say who speaks. Keep lines short so they fit inside 10 seconds.
  • On-screen text: keep it to a few words, put them in quotes and say where they appear. The API docs say text renders readably, while the model card still lists accurate text rendering as a hard case. Check every frame before you publish.
  • Language: the docs say English is fully supported; other languages have not been evaluated, so results may vary.

Length, cuts and extending a scene

Each generation is 3 to 10 seconds. For longer pieces, Gemini Omni 1.1 Flash adds scene extension: according to Google's 1.1 announcement, it reads up to 10 seconds of prior context and extends in 10-second steps up to 40 seconds. An extension prompt only needs the next beat:

  • "Continue the scene. She reaches the door and opens it."
  • Keep the same style and audio words you used in the first clip.
  • Extension appends to the end only, and the API docs note you can't extend an uploaded clip of someone talking to add new dialogue.

For draft work, the 1.1 release also adds a 360p draft mode that Google says is faster to generate. Lock your prompt at 360p, then render the keeper at 720p or higher.

Common mistakes and quick fixes

ProblemLikely causeFix
Unwanted cutsNo shot instructionAdd "in a single continuous shot, no scene cuts"
Music you didn't wantAudio left unspecifiedDescribe the audio or add "no music"
Edit changed too muchLong, multi-change edit promptOne change per turn, plus "keep everything else the same"
Character drifts between clipsText-only descriptionsAttach a reference image of the character
Garbled sign or captionLong or unquoted textShorter text in quotes; check the result
Rushed actionToo many events for 10 secondsCut to two or three beats, or extend the scene

Where to run these prompts

You can use these prompts anywhere Gemini Omni Flash is offered: the Gemini app and Google Flow for Google AI plan subscribers, Google AI Studio and the Gemini API for developers, and multi-model platforms. On Kenerate, Gemini Omni Flash sits in the AI video generator next to Veo 3.1, Kling v3 Pro, Seedance 2.5, Sora 2 Pro and Kenerate's own Ken 1, so you can run one prompt across several models and keep the best take. New accounts get 25 free credits, and credits never expire.

For the bigger picture, read what Gemini Omni Flash is and what it can do and our walkthrough of how to use Gemini Omni Flash in the app and API. If you start from stills, the image-to-video and reference-to-video guides cover how to prepare a clean reference, and the video upscaler helps when a draft needs a sharper final render.

Bottom line

Gemini Omni Flash rewards clear intent over long prompts. Describe the action, the shot and the sound; ask for a single continuous shot when you need one; and edit one small change at a time. Start from the six examples above, swap in your own subject, and save the versions that work as your house templates.