How to prompt Gemini Omni 1.1 Flash: short dialogue lines, ambient sound and vertical framing, with a reusable template, three example prompts and mistakes.
Gemini Omni 1.1 Flash, from Google, treats sound as part of the shot. Audio is always on, covering dialogue and ambience, so a good prompt reads like a short script page: who is in frame, what they do, what they say and what the place sounds like. This guide covers how the model reads its inputs, a template, three example prompts and the habits that help.
The prompt is required, and it can drive the shot on its own. You can also add a first frame, or a first and last frame, to fix the opening and ending images. Up to 6 reference images can join the prompt, and they combine with frames, which helps when a character from your references has to appear in a framed shot.
Clips run from 3 to 10 seconds at 360p, 720p, 1080p or 4K, in 16:9 or 9:16. Sound is always on, so it pays to write the audio yourself rather than leave it to chance.
Treat the prompt as three layers: the picture (framing, subject, action, light), the words (who says what, and how), and the sound bed (ambience and effects). Not every shot needs all three, but every shot gets sound.
[Shot size or camera move] of [characters] [action], [setting]. The [character] [delivery verb], '[short line]'. The [other character] [delivery verb], '[reply]'. [Light]. Sound: [ambience], [effects].
For a shot without dialogue, drop the lines and keep the Sound line. For vertical video, set 9:16 and start the prompt with "Vertical video:" so the prompt and the setting agree.
Made in Cinemagiq. The exact prompt is under each one.
PromptClose two-shot of two kids hiding under a blanket fort with a flashlight. The girl whispers, 'Did you hear that?' The boy whispers back, 'It's just the cat.' Warm flashlight glow.
PromptSlow pan across a rainy cafe terrace at night as a waiter stacks chairs under the awning. Sound: rain on the awning and distant traffic.
PromptVertical video: a skateboarder rolls toward the camera down an empty city street at sunrise and kicks the board up into his hand. Sound of wheels on asphalt.
Example 1 is a dialogue shot. It sets the framing first ("Close two-shot of two kids"), then gives each line a speaker and a delivery: "The girl whispers, 'Did you hear that?'" and "The boy whispers back, 'It's just the cat.'" Both lines are short enough for a 5-second clip, and "Warm flashlight glow" closes on the light.
Example 2 has no dialogue, so the sound carries the mood. The camera move comes first ("Slow pan across a rainy cafe terrace at night"), with a simple human action inside it ("a waiter stacks chairs under the awning"). The Sound line layers a close source and a far one: "rain on the awning and distant traffic".
Example 3 uses the 9:16 setting and says so: "Vertical video:". The action moves toward the lens ("rolls toward the camera"), which suits a tall frame, and it ends on a clear final beat, "kicks the board up into his hand". The sound is written inline, "Sound of wheels on asphalt", rather than as a separate line.
For shot sizes, camera moves and more sound cues, see the prompt engineering cheat sheet.
Yes. Sound is always on, and it covers dialogue and ambience. Put each line in quotes, say who speaks it and how, for example 'The girl whispers', and keep the lines short enough to fit a clip of 3 to 10 seconds.
A prompt alone, a first frame, or a first and last frame. You can also add up to 6 reference images, and they can be combined with frames in the same generation. The prompt is always required, even when frames or references set the look.
It renders 16:9 or 9:16, at 360p, 720p, 1080p or 4K. Set the ratio in the settings, then write the prompt for it. In vertical video, movement toward or away from the camera fills the frame better than wide sideways action.
From 3 to 10 seconds per generation. That is enough for one shot with a line or two of dialogue. For a longer scene, plan it as several shots and generate each one with its own prompt, then add each take to its shot on the Production Board.
Run Gemini Omni 1.1 Flash and the other leading models in one AI-native workspace.