--- title: Prompting Gemini Omni Flash 1.1 — Gemini Omni Flash 1.1 | Runware Docs url: https://runware.ai/docs/models/google-gemini-omni-flash-1-1/guides/prompting description: "How to prompt Gemini Omni Flash 1.1 for video: the five-element structure, camera vocabulary, holding a single shot, directing audio, and the sampling controls." --- ### [Introduction](https://runware.ai/docs/models/google-gemini-omni-flash-1-1/guides/prompting#introduction) Gemini Omni Flash 1.1 takes a **single text prompt** and returns a finished shot with its **soundtrack already in it**. There is no audio toggle and no second pass. Whatever you write about picture and sound lands in the same clip, which makes the prompt the only lever that matters for most of what you generate. The model is tuned to read a **five-element structure**: how the shot is framed and how the camera moves, the style, the light, the location, and the action. Three or four sentences covering those beats consistently outperform a paragraph of cinematic adjectives. [Watch video](https://runware.ai/docs/assets/output-hero.VxlASXhC.mp4) *One prompt, five elements, sound included. Play it with audio on.* > **Prompt**: A low tracking shot running alongside the subject at knee height. Commercial sportswear cinematography, crisp and high-contrast. Cold blue pre-dawn light with the first warm sun breaking along the horizon. An empty city waterfront promenade, wet concrete from overnight rain, the skyline across the water still lit. A female runner in a charcoal training top pushes through the last stretch of a session, breathing hard and steady, her shoes striking through shallow puddles. The audio is her rhythmic footfalls on wet concrete, controlled breathing, and a low harbour ambience underneath, no music. Play the clip with sound: the footfalls land on the strides. This guide covers the five-element structure, the camera vocabulary the model reads as direction, how to stop it from cutting, how to direct the audio, and the two sampling controls that shape how literally it reads you. ### [The request](https://runware.ai/docs/models/google-gemini-omni-flash-1-1/guides/prompting#the-request) A text-to-video call needs the prompt. Everything else has a working default. TypeScriptPythoncURLCLIJSON ```typescript import { createClient } from '@runware/sdk' const client = await createClient({ apiKey: process.env.RUNWARE_API_KEY }) await client.connect() const [result] = await client.run({ model: 'google:gemini@omni-flash-1.1', positivePrompt: 'A low tracking shot running alongside the subject at knee height. Commercial sportswear cinematography. Cold blue pre-dawn light with the first sun on the horizon. An empty city waterfront promenade, wet concrete from overnight rain. A female runner pushes through the last stretch of a session, her shoes striking through shallow puddles. The audio is her footfalls on wet concrete and controlled breathing, no music.', width: 1920, height: 1080, duration: 8 }) ``` ```python import asyncio import os from runware import Runware async def main(): async with Runware(api_key=os.environ["RUNWARE_API_KEY"]) as client: results = await client.run({ "model": "google:gemini@omni-flash-1.1", "positivePrompt": "A low tracking shot running alongside the subject at knee height. Commercial sportswear cinematography. Cold blue pre-dawn light with the first sun on the horizon. An empty city waterfront promenade, wet concrete from overnight rain. A female runner pushes through the last stretch of a session, her shoes striking through shallow puddles. The audio is her footfalls on wet concrete and controlled breathing, no music.", "width": 1920, "height": 1080, "duration": 8 }) asyncio.run(main()) ``` ```bash curl https://api.runware.ai/v1 \ -H "Authorization: Bearer $RUNWARE_API_KEY" \ -H "Content-Type: application/json" \ -d '[ { "taskType": "videoInference", "taskUUID": "b7c2d1e4-8f9a-4023-b1c2-d3e4f5061728", "model": "google:gemini@omni-flash-1.1", "positivePrompt": "A low tracking shot running alongside the subject at knee height. Commercial sportswear cinematography. Cold blue pre-dawn light with the first sun on the horizon. An empty city waterfront promenade, wet concrete from overnight rain. A female runner pushes through the last stretch of a session, her shoes striking through shallow puddles. The audio is her footfalls on wet concrete and controlled breathing, no music.", "width": 1920, "height": 1080, "duration": 8 } ]' ``` ```bash runware run google:gemini@omni-flash-1.1 \ positivePrompt="A low tracking shot running alongside the subject at knee height. Commercial sportswear cinematography. Cold blue pre-dawn light with the first sun on the horizon. An empty city waterfront promenade, wet concrete from overnight rain. A female runner pushes through the last stretch of a session, her shoes striking through shallow puddles. The audio is her footfalls on wet concrete and controlled breathing, no music." \ width=1920 \ height=1080 \ duration=8 ``` ```json { "taskType": "videoInference", "taskUUID": "b7c2d1e4-8f9a-4023-b1c2-d3e4f5061728", "model": "google:gemini@omni-flash-1.1", "positivePrompt": "A low tracking shot running alongside the subject at knee height. Commercial sportswear cinematography. Cold blue pre-dawn light with the first sun on the horizon. An empty city waterfront promenade, wet concrete from overnight rain. A female runner pushes through the last stretch of a session, her shoes striking through shallow puddles. The audio is her footfalls on wet concrete and controlled breathing, no music.", "width": 1920, "height": 1080, "duration": 8 } ``` Response ```json [ { "taskType": "videoInference", "taskUUID": "b7c2d1e4-8f9a-4023-b1c2-d3e4f5061728", "videoUUID": "4a5b6c7d-8e9f-4012-a3b4-c5d6e7f80912", "videoURL": "https://vm.runware.ai/video/os/a14d18/ws/2/vi/4a5b6c7d-8e9f-4012-a3b4-c5d6e7f80912.mp4" } ] ``` - `positivePrompt` is the only required field. It accepts up to **40000 characters**, which is far more headroom than a good prompt needs. - `width` and `height` are optional and travel as a pair. `resolution` is the alternative way to size the output. Both are covered in the [resolution and duration guide](https://runware.ai/docs/models/google-gemini-omni-flash-1-1/guides/resolution-and-duration). - `duration` is a whole number of seconds from **3 to 10**, defaulting to **6**. - `settings.temperature` and `settings.topP` shape how far the model strays from the most obvious reading of your prompt. See [Sampling controls](https://runware.ai/docs/models/google-gemini-omni-flash-1-1/guides/prompting#sampling-controls). Audio is generated natively on every call. **Attaching media to `inputs` switches the model into one of its other modes**, covered in the [reference](https://runware.ai/docs/models/google-gemini-omni-flash-1-1/guides/reference-driven-video), [frame](https://runware.ai/docs/models/google-gemini-omni-flash-1-1/guides/first-and-last-frame), [editing](https://runware.ai/docs/models/google-gemini-omni-flash-1-1/guides/editing-video), and [extension](https://runware.ai/docs/models/google-gemini-omni-flash-1-1/guides/extending-video) guides. The prompting principles here carry into all of them. ### [The five-element prompt structure](https://runware.ai/docs/models/google-gemini-omni-flash-1-1/guides/prompting#the-five-element-prompt-structure) The five elements are **shot framing and motion, style, lighting, location, and action**. The hero above is built from exactly those beats, in that order. Hover a segment to see which element it carries: **[Shot framing and motion]** A low tracking shot running alongside the subject at knee height, **[Style]** commercial sportswear cinematography, crisp and high-contrast, **[Lighting]** cold blue pre-dawn light with the first warm sun breaking along the horizon, **[Location]** an empty city waterfront promenade, wet concrete from overnight rain, the skyline across the water still lit, **[Action]** a female runner in a charcoal training top pushes through the last stretch of a session, her shoes striking through shallow puddles Every element names a **decision a director would still have to make**. None of them describe what the model already knows. Wet concrete reflects, pre-dawn light is cold, a hard session means visible effort. Spending prompt on those facts buys nothing. **Drop an element when the shot genuinely doesn't need it.** A flat-lit interior needs no lighting clause, a tight portrait needs no location. Treat the five as a checklist against forgetting one, not as a template to fill. ### [Camera language](https://runware.ai/docs/models/google-gemini-omni-flash-1-1/guides/prompting#camera-language) The model reads **cinematic camera vocabulary as instruction** rather than as scene description. Naming a move gets you that move. Describing the effect of a move gets you a guess. The same apartment below, filmed three ways for a property listing. The scene is fixed and only the camera grammar changes. **Oner**: [Watch video](https://runware.ai/docs/assets/output-camera-oner.BX98UHG2.mp4) *A continuous oner through the whole flat* > **Prompt**: A continuous oner moving through a newly finished two-bedroom apartment, in a single unbroken scene with no cuts. The camera glides from the entry hall past the open kitchen, turns through the living room, and settles facing the balcony doors. Bright late-morning daylight through full-height windows, pale oak floors, a linen sofa, and a marble kitchen island. Real-estate listing cinematography, steady and even. Soft ambient room tone, no music, no dialogue. **Push in**: [Watch video](https://runware.ai/docs/assets/output-camera-pushin.kaC1Hrgz.mp4) *A slow push in onto the view* > **Prompt**: A slow steady push in toward the balcony doors of the same newly finished apartment, in a single unbroken scene with no cuts. The camera starts wide in the living room and moves forward at a constant pace until the city view fills the frame. Bright late-morning daylight, pale oak floors, a linen sofa, a marble kitchen island at the edge of frame. Real-estate listing cinematography. Soft ambient room tone, no music, no dialogue. The oner flows room to room without a break. The push in holds one axis and advances at a constant rate. **Each phrase is an instruction to the operator**, not a description of the result. The third phrase changes the shape of the shot as well as the move. A handheld smartphone tour brings the hand wobble and the looser amateur read a listing story wants, and it is **the one you would frame vertically**. [Watch video](https://runware.ai/docs/assets/output-camera-smartphone.QXYNYxg5.mp4) *A handheld phone tour, framed 720 × 1280 for a listing story* > **Prompt**: A natural handheld smartphone tour of the same newly finished apartment, shot vertically as an agent would film it for a listing story. The camera walks from the entry hall into the living room with the gentle organic wobble of a phone held in one hand, panning across the kitchen island and up to the full-height windows. Bright late-morning daylight, pale oak floors, a linen sofa. Soft ambient room tone, no music, no dialogue. **Lead with the camera phrase.** The model treats the opening words as the load-bearing statement of how to film the shot. A move named in the third sentence tends to be absorbed as scene description instead. ### [Holding a single shot](https://runware.ai/docs/models/google-gemini-omni-flash-1-1/guides/prompting#holding-a-single-shot) The model **defaults to cutting**. Left to itself it will treat a prompt as a small sequence and assemble two or three shots, which is useful for a reel and wrong for anything that has to read as one continuous take. The instruction that stops it is explicit: **"in a single unbroken scene"** or **"no scene cuts"**. The pair below is the same mountain-pass prompt with and without that clause. **Default**: [Watch video](https://runware.ai/docs/assets/output-scene-default.BryYbLI5.mp4) *Default: the model assembles cuts* > **Prompt**: An electric SUV drives a high mountain pass at golden hour. Automotive commercial cinematography, deep saturated grade. Warm low sun raking across bare rock and a guardrail, long shadows across the asphalt. The car takes a sweeping bend with the valley falling away beside it. The audio is tyre noise on dry asphalt, a faint electric whine, and open wind, no music. **Single scene**: [Watch video](https://runware.ai/docs/assets/output-scene-single.d73x_3Lm.mp4) *With the single-scene clause: one continuous take* > **Prompt**: An electric SUV drives a high mountain pass at golden hour, in a single unbroken scene with no scene cuts. One continuous aerial shot tracking the car from the side as it takes a sweeping bend with the valley falling away beside it. Automotive commercial cinematography, deep saturated grade. Warm low sun raking across bare rock and a guardrail, long shadows across the asphalt. The audio is tyre noise on dry asphalt, a faint electric whine, and open wind, no music. The default clip covers the same ground in cuts. The single-scene clip **holds one camera for the full runtime**, which is what a hero banner or a looping background needs. > [!NOTE] > Decide this before you write anything else, because it changes how the rest of the prompt should read. A cutting prompt describes **beats** ("begin on the wheels, cut to the driver"). A single-scene prompt describes **one continuous move** ("one continuous aerial tracking the car through the bend"). Mixing the two grammars gives the model contradictory instructions. ### [Directing the audio](https://runware.ai/docs/models/google-gemini-omni-flash-1-1/guides/prompting#directing-the-audio) Sound comes from the same prompt as the picture, so **write the audio as direction** and not as an afterthought. Three patterns cover most needs: name each foley element, name the layers of an ambient bed, or quote the line you want spoken. [Watch video](https://runware.ai/docs/assets/output-audio-foley.C1i-i1KC.mp4) *Named foley: every sound listed by hand* > **Prompt**: A close product shot of hands typing on a mechanical keyboard at a clean desk setup, in a single unbroken scene. Even daylight from the left, a matte dark desk mat, a monitor softly defocused behind. Tech product cinematography, shallow depth of field. The audio foregrounds precise foley: the sharp individual clack of each keypress, the deeper thunk of the spacebar, the light scrape of a fingertip across a keycap edge, and a quiet room tone underneath. No music, no dialogue. [Watch video](https://runware.ai/docs/assets/output-audio-ambient.BAsskD8x.mp4) *Layered ambience: the bed named layer by layer* > **Prompt**: A slow wide shot across a beachfront hotel terrace in the late afternoon, in a single unbroken scene. Rattan loungers under cream parasols, a long infinity pool catching the low sun, the sea beyond. Hospitality brand cinematography, warm and airy. The audio is a layered ambient bed: the steady wash of surf below the terrace, a light breeze moving through the parasol fabric, the occasional soft clink of glassware from an unseen bar, and distant gulls. No music, no dialogue. [Watch video](https://runware.ai/docs/assets/output-audio-dialogue.Cxt07xV1.mp4) *Quoted dialogue: the exact line, in a named accent* > **Prompt**: A clean medium close-up of a product marketing manager in her thirties speaking directly to camera in a bright modern office, in a single unbroken scene. She says, in a warm and measured British accent: "We rebuilt the dashboard around one question. What do you need to see first?" Her lips and expression stay in sync with every word. Soft key light from a window on the left, a blurred workspace behind her. The audio is her clear spoken voice up front over a quiet office room tone. No music. The keyboard prompt names **each foley element separately**, and the model builds that stack instead of a generic typing sound. The terrace prompt names the layers of the bed, and the balance holds with no single layer swallowing the others. The spokesperson prompt **quotes the line and names the accent**, so the words arrive as written rather than improvised around the topic. **Say what you don't want too.** Left unsaid, the model tends to score a cinematic shot with music of its own. Adding "no music" keeps the track clean for a clip that will sit under a soundtrack you control, and "no dialogue" stops invented voiceover on a shot with a person in it. ### [Sampling controls](https://runware.ai/docs/models/google-gemini-omni-flash-1-1/guides/prompting#sampling-controls) Two settings change how literally the model reads you. `settings.temperature` runs from **0 to 2** and defaults to 1, so the default sits in the middle of the range. Turn it **down toward 0** to hold the model to the most obvious reading of the prompt, or **up toward 2** to loosen it for more variation between takes. `settings.topP` also runs from 0 to 1, defaults to 0.95, and **narrows the pool of candidates** the model samples from at each step. **temperature 0.2**: [Watch video](https://runware.ai/docs/assets/output-temperature-low.jZOr-412.mp4) *temperature 0.2: the literal read* > **Prompt**: A slow orbit around a tall glass of iced citrus drink on a pale stone counter, in a single unbroken scene. Beverage brand cinematography, bright and clean. Hard midday sun through a window casting a sharp shadow across the counter, condensation beading on the glass, a slice of blood orange on the rim. The audio is ice settling in the glass and a quiet kitchen room tone, no music. **Default**: [Watch video](https://runware.ai/docs/assets/output-temperature-default.M12uVUiD.mp4) *Default: the same prompt, given room* > **Prompt**: A slow orbit around a tall glass of iced citrus drink on a pale stone counter, in a single unbroken scene. Beverage brand cinematography, bright and clean. Hard midday sun through a window casting a sharp shadow across the counter, condensation beading on the glass, a slice of blood orange on the rim. The audio is ice settling in the glass and a quiet kitchen room tone, no music. Both clips ran the same prompt and **differ only in temperature**. The low-temperature take stays with the plain reading of every clause. The default take keeps the brief but makes freer choices about staging and styling around it. **Turn the temperature down when the brief is fixed**, which is most commercial work: a named product, an approved look, a shot that has to match a board. **Leave it at the default, or raise it toward 2, while exploring**, when you want several distinct takes of a loose idea to choose from. Leave `topP` alone unless output feels repetitive across takes at a temperature you otherwise like. ### [Tips](https://runware.ai/docs/models/google-gemini-omni-flash-1-1/guides/prompting#tips) 1. **Cover the five elements once each.** Shot framing and motion, style, lighting, location, action. A missing element gets defaulted to whatever the rest of the prompt implies, which is often not the choice you would have made. 2. **Open with the camera.** "A continuous oner moving through...", "A slow steady push in toward...". The first clause sets the grammar for everything after it. 3. **Say "in a single unbroken scene" whenever you need one take.** The model cuts by default. This is the single most common surprise in its output and the fix is one clause. 4. **Write the audio as direction.** Name each foley element, name the layers of an ambient bed, or quote the spoken line and its accent. Silence about sound gets you the model's own choice, which usually includes music. 5. **Add "no music" and "no dialogue" when you mean it.** Both defaults lean the other way, and both are negated by naming the exclusion. 6. **Don't describe what the model already knows.** Wet asphalt reflects, low sun is warm, a hard run means visible effort. Prompt the decisions, not the physics. 7. **Drop the temperature for brief-bound work.** A named product or an approved look wants the literal read, so turn temperature toward 0. The default of 1 is the middle of the 0 to 2 range, and raising it loosens the read for exploration. 8. **Keep it to three or four sentences.** That covers the five elements without drifting into specification. Longer prompts are harder to iterate and rarely better.