Gemini Omni 1.2 Flash

Gemini Omni 1.2 Flash is Google’s video generation and editing model. It makes videos with a synchronized audio track from a text prompt, images, or an existing video. Clips run from 3 to 15 seconds.

Compared with Gemini Omni 1.1, version 1.2 makes clips up to 15 seconds instead of 10, and you can set the length to the second. Google also reports that it holds onto faces, clothing, product labels and brand colors more closely across a clip, and that dialogue sounds more natural with better lip sync.

What you can do

• Generate a video from a text prompt.
• Animate a starting image with image.
• Set a first and a last frame with image and last_frame, and the model fills in the motion between them. Interpolation supports up to 10 seconds.
• Pass up to 6 reference_images of a character, product or style. The model uses them to keep the subject consistent, not as literal frames.
• Edit an existing video of up to 15 seconds with video. Describe the change in the prompt, such as “make it night and raining, keep everything else the same”, and the model leaves the rest of the clip alone.
• Choose 360p for fast drafts, 720p, or 1080p and 4k, which Google upscales from the base render. Output is 16:9 or 9:16 at 24 fps.

Inputs that work together

image and last_frame can’t be combined with reference_images, and video can’t be combined with any image input. duration and aspect_ratio don’t apply to edits, because the output keeps the length and framing of the source video.

Usage

Text to video

import replicate

output = replicate.run(
    "google/gemini-omni-1.2-flash",
    input={
        "prompt": "A single unbroken tracking shot following a paper airplane as it glides through a busy library. Pages rustle and a librarian whispers 'hey!' as it passes. No music.",
        "duration": 15,
        "resolution": "720p",
    },
)
print(output)

Reference images

output = replicate.run(
    "google/gemini-omni-1.2-flash",
    input={
        "prompt": "The woman from the first image walks through a farmers market holding the tumbler from the second image and takes a sip.",
        "reference_images": [
            "https://example.com/character.png",
            "https://example.com/product.png",
        ],
        "aspect_ratio": "9:16",
        "duration": 10,
    },
)

Edit a video

output = replicate.run(
    "google/gemini-omni-1.2-flash",
    input={
        "prompt": "Change the scene to night in heavy rain. Keep the people, their dialogue and the camera move the same.",
        "video": "https://example.com/clip.mp4",
    },
)

Prompting tips

• Describe the shot framing, camera move, lighting, setting and action. More detail gives you more control.
• Write the audio you want. Put dialogue in quotes and name who says it, and describe ambient sound and music. Write “no music” or “no dialogue” to leave them out.
• Ask for “one continuous shot” to avoid cuts. Camera terms like “push in”, “orbit”, “dolly zoom” and “FPV drone” work.
• The duration input wins over any length you write in the prompt.
• For edits, say what should stay the same as well as what should change.
• Render drafts at 360p, then rerun the prompt you like at a higher resolution.

Notes

• Every video carries an invisible SynthID watermark.
• Google’s safety filters apply to prompts, input media and generated video. A prompt or output they block fails with a content error.
• Google’s announcement: Gemini Omni 1.2 Flash. Prompt guide: Gemini Omni prompting.